ultralytics

🌟 Summary

Ultralytics v8.4.118 introduces a standalone OpenAI-compatible LLM interface alongside YOLO, while improving OBB training, dataset handling, model training reliability, and documentation workflows. 🚀

📊 Key Changes

• 🤖 New standalone LLM model interface by @glenn-jocher

  • Add from ultralytics import LLM for text and image-based language-model requests.
  • Supports OpenAI Responses and Chat Completions APIs, including synchronous and asynchronous calls.
  • Accepts images from local paths, URLs, data URLs, NumPy arrays, and PIL images.
  • Supports reusable prompts, request overrides, conversation state, API keys, and OpenAI-compatible service endpoints.
  • Uses the optional openai dependency and remains independent of Ultralytics Platform and workflow-runtime components.

• 📐 Improved oriented bounding box training

  • Mosaic, CutMix, and RandomPerspective now preserve OBB orientation when objects are clipped by image boundaries.
  • Prevents clipped objects from receiving incorrect rotation angles during training.

• ⚡ Faster CopyPaste augmentation

  • Batches instance concatenation instead of repeatedly copying growing arrays.
  • Reduces unnecessary processing overhead, especially when many objects are copied.

• 🧠 More reliable YOLOE behavior

  • Validates visual prompts before modifying model state.
  • Accepts flat prompts for supported batched image sources.
  • Rejects invalid string class labels and mismatched vocabularies earlier with clearer errors.
  • Preserves gradient settings when converting YOLOE convolution layers to linear layers.

• 🏋️ Training and inference stability fixes

  • Correctly resets dataloader workers when resuming after Mosaic augmentation is closed.
  • Allows repeated train() and tune() calls on the same model object.
  • Prevents duplicate World model callbacks across multi-dataset training.
  • Fixes classification prediction for models without predefined transforms.
  • Ensures classification validation loaders do not discard samples when compiling.
  • Fixes SAM and related predictor models being created with incompatible inference-only tensors.

• 🗂️ Dataset and prediction improvements

  • Classification auto-splitting now recognizes all supported image formats, including JPEG, BMP, WebP, TIFF, AVIF, HEIC, and uppercase extensions.
  • Missing classification images now raise a clear FileNotFoundError instead of failing later with an unrelated directory error.
  • Preserves original filenames when loading images after EXIF correction.
  • Keeps bounding-box fallbacks for malformed grounding segmentation labels.

• 📚 Documentation and deployment updates

  • Standardizes strict documentation validation on Zensical and updates contributor instructions.
  • Documentation redeployment now detects Python docstring and all configuration-file changes.
  • Restores model benchmark chart placeholders, including for YOLO26, while moving production site features to the centralized publisher.
  • Documents replacing the model behind an existing deployment without changing its endpoint URL, API key, or deployment identity.
  • Updates Albumentations examples for current 2.x constructor names and refreshes the Rust inference dependency to 0.0.34.
  • Adds API reference documentation for the new LLM interface.

🎯 Purpose & Impact

• 🚀 Broader AI capabilities: Developers can now use Ultralytics as a unified entry point for YOLO vision models and OpenAI-compatible language models, including multimodal image understanding.
• 🔌 Flexible integration: The new LLM class works with OpenAI and compatible providers without requiring Platform or workflow features.
• 🎯 Better OBB accuracy: Rotated-object datasets should receive more consistent training targets when augmentation crops objects at image edges.
• ⚡ Improved performance: CopyPaste augmentation can run more efficiently, particularly on images containing many instances.
• 🛡️ More predictable training: Resume, compile, repeated training, World models, SAM, and classification workflows are less likely to produce silent errors or invalid results.
• 🧰 Easier maintenance: Broader dataset format support, clearer validation errors, and more accurate documentation reduce setup and debugging time for users.

What's Changed

• Preserve OBB orientation through clipped augmentations by @Nikhi00718 in https://github.com/ultralytics/ultralytics/pull/25723
• Redeploy docs when Python or cfg sources change by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25733
• Standardize Docs validation on Zensical by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25736
• Simplify strict Docs validation and restore model charts by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25738
• Document deployment model replacement by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25745
• Remove the YOLO11 podcast audio embed by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/25765
• Align grounding segmentation fallback by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/25743
• Batch the concatenation loop in CopyPaste.apply_instances to avoid O(n^2) growth by @JESUSROYETH in https://github.com/ultralytics/ultralytics/pull/25732
• Update the Albumentations examples to the 2.x constructor arguments by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/25759
• Standardize export data defaults by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/25742
• Cache external test assets before pytest by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25737
• Accept all image formats in classification auto-split and report missing images instead of crashing by @doublecurry in https://github.com/ultralytics/ultralytics/pull/25749
• Validate yoloe visual prompts by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/25744
• Reject a short YOLOE vocabulary before the head is re-parameterized by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/25747
• Build loaded models outside inference mode by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/25748
• Register the World pretrain hook once per model by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/25758
• Build the name=model save directory from the model stem by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/25763
• Reset dataloader workers when resuming after mosaic closure by @JESUSROYETH in https://github.com/ultralytics/ultralytics/pull/25762
• Restore the model override after training by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/25760
• Fix: autocast_list loses image filename after exif_transpose by @Q-qqq in https://github.com/ultralytics/ultralytics/pull/25741
• Fix classification predict crash when the model has no transforms by @ahmet-f-gumustas in https://github.com/ultralytics/ultralytics/pull/25735
• Build SAM predictor models outside inference mode by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/25756
• fix: only apply drop_last to the classify train loader under compile, not val by @JESUSROYETH in https://github.com/ultralytics/ultralytics/pull/25734
• Revert "Build loaded models outside inference mode" by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25768
• Name the ImageNet pseudo-label teacher checkpoints and correct stored dtype by @Bovey0809 in https://github.com/ultralytics/ultralytics/pull/25754
• Bump ultralytics-inference docs version to 0.0.34 by @onuralpszr in https://github.com/ultralytics/ultralytics/pull/25778
• Add standalone LLM model interface by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25761

New Contributors

• @doublecurry made their first contribution in https://github.com/ultralytics/ultralytics/pull/25749

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.117...v8.4.118

🌟 Summary

v8.4.117 improves augmentation correctness, model reliability, deployment safety, and documentation across Ultralytics YOLO and YOLO26. 🚀

📊 Key Changes

• 🧩 Albumentations now handles spatial transforms by type

  • PR #25633 replaces the fragile hardcoded transform-name list with recursive type detection, so wrapped transforms such as OneOf correctly update annotations.
  • Spatial augmentations now carry bounding boxes, polygons, masks, depth maps, and keypoints through the pipeline.
  • Background-only images receive pixel-level augmentations even when they have no annotations.
  • Keypoint flipping respects the configured flip_idx mapping.
  • Probability handling now makes a zero probability a reliable off switch.
  • Unsupported topology-changing transforms, such as grid shuffling with polygons or keypoints, are detected instead of silently corrupting labels.

• 🔐 Improved security for dependency installation

  • check_requirements() now prevents untrusted requirement strings from being interpreted as shell commands.
  • This protects workflows that automatically install missing dependencies while loading model files.

• 🛡️ More reliable dataset and mask processing

  • Fixed Windows semantic-mask shape handling so grayscale masks remain two-dimensional and semantic training no longer fails during Mosaic augmentation.
  • COCO conversion now keeps one consistent label format per file, warns about unusable polygons, and falls back to box-shaped polygons when necessary.
  • Mixed detection and segmentation rows are rejected instead of being silently misread.
  • Grounding dataset caching and validation were improved, including clearer errors for empty or invalid annotation sets.
  • Degenerate polygon contours are skipped when exporting labels.

• 🧠 Depth estimation improvements

  • Depth postprocessing now aligns PyTorch, Hailo, and exported-model outputs before removing padding, producing more consistent results across inference backends.
  • Depth validation no longer inherits rectangular batching behavior that conflicts with its intentional square stretching.
  • The KITTI depth configuration now uses the canonical 652-frame left-camera evaluation split, avoiding test-set overlap and making reported results more reproducible.
  • Depth metrics and documentation were updated to reflect per-image evaluation behavior.

• ⚡ Faster and more consistent inference

  • YOLO26 end-to-end postprocessing uses grouped top-k selection, improving TensorRT FP16 latency by approximately 1.8% to 8.1% without changing mAP.
  • RT-DETR FLOPs profiling is faster and now supports attention-based decoder architectures correctly.
  • SavedModel INT8 calibration avoids an unnecessary NumPy copy.
  • GMC tracking now handles textureless frames and incomplete feature matches without stopping, allowing tracking to recover cleanly.
  • SAM3 video masks use the model’s configured threshold consistently instead of a fixed threshold.

• 🎯 Expanded model and training support

  • Pose training now accepts dataset-defined kpt_oks_sigmas, with validation that the configuration matches the model’s keypoint count.
  • OBB utilities handle empty inputs more safely, and documentation clarifies long-edge canonicalization and prediction formats.
  • Predictors now report unsupported options such as augmentation, embeddings, and visualization instead of silently ignoring them.
  • Cached predictors are refreshed after training so predictions use the newly trained weights and class names.
  • Embedding requests now provide clearer errors for exported or third-party models that do not expose compatible internal layers.
  • YOLOE class and visual-prompt validation was strengthened, supporting more reliable prompt-free vocabulary workflows.

• 📚 Documentation and platform updates

  • Export documentation now covers additional options such as name, split, conf, iou, max_det, and agnostic_nms.
  • Classification dataset inputs are consistently documented as directories or built-in dataset names rather than YAML files.
  • Added reusable depth-speed comparison documentation and a new YAML2ModelGraph integration guide for generating YOLO architecture diagrams from model YAML files.
  • Added Platform troubleshooting guidance for datasets, training, deployment, billing, and common questions.
  • FAQ and supporting sections were reorganized for more consistent documentation rendering.
  • Rust inference documentation now references ultralytics-inference version 0.0.33.
  • Documentation publishing is now restricted to the main branch to prevent accidental production releases.

🎯 Purpose & Impact

• ✅ More trustworthy augmentation: Labels and auxiliary data remain aligned when using custom or nested Albumentations pipelines, reducing silent training errors.
• 🧪 Better training stability: Depth, semantic segmentation, pose, OBB, grounding, and tracking workflows handle edge cases more gracefully.
• 🚀 Improved deployment performance: YOLO26 TensorRT exports can achieve lower postprocessing latency, while depth predictions behave more consistently across backends.
• 🔒 Safer model loading: Dependency checks no longer expose shell execution risks through malicious requirement strings.
• 📊 More reproducible evaluation: The corrected KITTI split and clarified depth metrics make comparisons easier to interpret, although results from older downloaded datasets may require rebuilding.
• 📖 Clearer user experience: Updated documentation explains task-specific inputs, export controls, model limitations, and platform troubleshooting in more practical terms.

What's Changed

• Bump eviden-actions/clean-self-hosted-runner from v1.4.35 to v1.4.36 in /.github/workflows by @UltralyticsAssistant in https://github.com/ultralytics/ultralytics/pull/25711
• Bump ultralytics-inference version to 0.0.33 in documentation by @onuralpszr in https://github.com/ultralytics/ultralytics/pull/25712
• Publish docs only from main by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25713
• Move the Depth Anything V2 speed table into a docs macro by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/25715
• Fix Windows semantic mask shape regression by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/25721
• Document the predictor data attribute as the args copy it holds by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/25692
• Threshold SAM3 video semantic masks at the model logit threshold by @JESUSROYETH in https://github.com/ultralytics/ultralytics/pull/25643
• Fix shell injection in check_requirements() via untrusted requirement strings by @Zenka737 in https://github.com/ultralytics/ultralytics/pull/25720
• Document max_det, iou, conf export arguments by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/25675
• feat: clarify confusion matrix confidence threshold during validation by @Rahulbiradar9 in https://github.com/ultralytics/ultralytics/pull/25677
• fix GMC empty descriptor handling by @Nikhi00718 in https://github.com/ultralytics/ultralytics/pull/25709
• Fix redundant SavedModel INT8 calibration copy by @amanharshx in https://github.com/ultralytics/ultralytics/pull/25639
• Fix depth val silently overriding its own stretch letterbox with inherited rect=True by @JESUSROYETH in https://github.com/ultralytics/ultralytics/pull/25646
• Document the OBB long-edge canonicalization, convert empty box inputs, and skip degenerate segment contours by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/25655
• Use the Hailo Model Zoo quantization recipe for YOLO26 HEF export by @nivosco in https://github.com/ultralytics/ultralytics/pull/25687
• Document the confidence threshold used for the val confusion matrix (#25674) by @Parth1353 in https://github.com/ultralytics/ultralytics/pull/25722
• Write one label format per file in convert_coco and document the classification data input by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/25672
• Grouped topk e2e postprocess by @artest08 in https://github.com/ultralytics/ultralytics/pull/25666
• Speed up RT-DETR FLOPs profiling by @Daniiiil1 in https://github.com/ultralytics/ultralytics/pull/25652
• Document missing export and predict arguments and the classification data directory by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/25690
• Support optional kpt_oks_sigmas in v8PoseLoss for training by @cosmo-gb in https://github.com/ultralytics/ultralytics/pull/25656
• Name the classification input in the export calibration, benchmark and tuner data rows by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/25698
• Report unimplemented predict arguments, refresh the predictor after training, and register tracker callbacks once per model by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/25688
• Remove brittle Hailo model-specific recipe by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25728
• Move documentation FAQs to page ends by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25730
• Use Ultralytics YOLO in task page titles and lead copy by @miles-deans-ultralytics in https://github.com/ultralytics/ultralytics/pull/25705
• Align depth postprocessing across inference backends by @JESUSROYETH in https://github.com/ultralytics/ultralytics/pull/25628
• Validate YOLOE classes and enable prompt-free vocabulary workflows by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/25691
• Fix grounding label scanning, caching, and reporting by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/25678
• Fix the KITTI Eigen test split: 3 test drives were in train by @JESUSROYETH in https://github.com/ultralytics/ultralytics/pull/25650
• Fix unbounded per-track solution state in continuous streams by @JESUSROYETH in https://github.com/ultralytics/ultralytics/pull/25729
• docs: add YAML2ModelGraph integration documentation by @Rahulbiradar9 in https://github.com/ultralytics/ultralytics/pull/25726
• Average DepthMetrics per image instead of pooling every valid pixel by @JESUSROYETH in https://github.com/ultralytics/ultralytics/pull/25648
• Gate DepthLoss26 gradient pyramids per image by @JESUSROYETH in https://github.com/ultralytics/ultralytics/pull/25637
• Route spatial Albumentations by type and carry masks, polygons and keypoints through it by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/25633

New Contributors

• @Parth1353 made their first contribution in https://github.com/ultralytics/ultralytics/pull/25722
• @Nikhi00718 made their first contribution in https://github.com/ultralytics/ultralytics/pull/25709
• @nivosco made their first contribution in https://github.com/ultralytics/ultralytics/pull/25687
• @Daniiiil1 made their first contribution in https://github.com/ultralytics/ultralytics/pull/25652

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.116...v8.4.117

🌟 Summary

🚀 v8.4.116 improves installation reliability, expands YOLOE and Platform workflows, strengthens tracking and export support, and refreshes YOLO26 documentation.

📊 Key Changes

• 🔧 OpenCV compatibility fix — current PR #25702 by @Y-T-G

  • Raises the minimum opencv-python version from 4.6.0 to 4.7.0.
  • Keeps the exclusion for 4.13.0.90, which is affected by a FIPS self-test crash.
  • Removes an outdated ONNX DNN backend requirement check.
  • This aligns the dependency with cv2.imdecodemulti, which Ultralytics uses internally.

• 🧠 Reusable YOLOE prompt embeddings

  • Adds save_prompt_embeddings() and load_prompt_embeddings() for storing text or visual prompt configurations in NPZ files.
  • Profiles are validated against the source YOLOE model and can be reused before exporting to formats such as ONNX, OpenVINO, TensorRT, CoreML, LiteRT, and RKNN.
  • Exported models remain standard single-input models and do not require the NPZ file at runtime.

• 📚 Improved model guidance

  • Reworks the model index into a task-and-mode comparison table.
  • Positions YOLO26 as the recommended model for new projects, with YOLO11 as a mature production alternative.
  • Clarifies support for YOLO12, OBB, SAM models, YOLOE, YOLO-World, RT-DETR, and other model families.
  • Adds a YOLO26 custom-dataset training video and highlights monocular depth estimation.

• 🎯 Broader and safer tracking support

  • Documents and supports OBB tracking alongside detection, segmentation, and pose.
  • Rejects unsupported semantic and depth tracking tasks with a clear error before processing begins.
  • Skips unnecessary camera-motion compensation work when gmc_method: none.
  • Keeps OC-SORT observation history bounded on all track lifecycle paths.

• 📦 More efficient model export

  • Streams ONNX and QNN calibration data instead of retaining all transformed images in memory.
  • Reduces calibration memory usage substantially for large datasets.
  • Updates anchor creation to use CoreML-friendly tensor operations, improving dynamic CoreML export compatibility.

• 🧪 Depth and segmentation fixes

  • Excludes ground-truth depth values outside the configured valid range during calibration, keeping calibration consistent with validation metrics.
  • Fixes FP16 segmentation with class-agnostic NMS.
  • Preserves YOLOE one-to-one classifier weights during linear probing, preventing a severe accuracy drop.
  • Makes pose activation-map gradients compatible with autograd and torch.compile.

• 🖼️ Visualization and analytics improvements

  • Restores percentage labels in analytics pie charts.
  • Speeds up semantic-mask overlay rendering by replacing repeated full-image scans with a palette lookup.

• ☁️ Expanded Ultralytics Platform workflows

  • Adds documented custom metadata support for datasets, images, projects, and models.
  • Supports nested metadata, metadata search, NDJSON image metadata, and Dataset Ingest API uploads.
  • Refreshes Platform integration screenshots and documents native Platform support for YOLOv8 and YOLOv5.
  • Refactors Platform callbacks to load only where needed, reducing unnecessary imports in prediction, validation, and export paths.

• 🛡️ Reliability and infrastructure

  • Makes downloads atomic, preventing concurrent test or application processes from reading partially written files.
  • Updates CI runner images to Node.js 24 for compatibility with newer tooling.
  • Refreshes OpenVINO benchmark data across Intel CPUs and NPUs.
  • Improves documentation deployment detection when configuration defaults change.

🎯 Purpose & Impact

• ✅ Fewer installation failures: Users relying on OpenCV image decoding now receive a compatible version automatically.
• 🚀 Simpler YOLOE deployment: Prompt configurations can be prepared once and reused across multiple export targets.
• 💾 Lower memory usage: Large ONNX and QNN calibration jobs are more practical, especially on limited-memory systems.
• 🎥 More capable tracking: OBB tracking is now clearly supported, while unsupported tasks fail with actionable messages instead of obscure runtime errors.
• 📈 Better model fine-tuning: YOLOE linear probing and depth calibration now preserve pretrained performance more reliably.
• 🧩 Improved deployment compatibility: Dynamic CoreML exports and FP16 segmentation workflows are more robust.
• 🔍 Better Platform organization: Custom metadata helps teams track provenance, review status, equipment, projects, and deployment context.
• 📖 Clearer onboarding: The refreshed model documentation makes it easier to choose the right Ultralytics model and understand its supported modes.

What's Changed

• Fix analytics pie chart percentage labels by @rudrakumar07 in https://github.com/ultralytics/ultralytics/pull/25600
• Prevent concurrent download cache corruption by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25612
• Fix dynamic CoreML anchor export by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25619
• Refactor Ultralytics Platform integration by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25622
• Fix FP16 end-to-end segmentation with class-agnostic NMS by @JESUSROYETH in https://github.com/ultralytics/ultralytics/pull/25613
• Preserve YOLOE one-to-one classifiers during linear probing by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/25607
• Name the PyPI version check step by @MGPOCKY in https://github.com/ultralytics/ultralytics/pull/25606
• Stream ONNX and QNN calibration data by @amanharshx in https://github.com/ultralytics/ultralytics/pull/25617
• Exclude out-of-range ground truth from depth calibration by @JESUSROYETH in https://github.com/ultralytics/ultralytics/pull/25614
• Build semantic overlays in a single pass by @JESUSROYETH in https://github.com/ultralytics/ultralytics/pull/25585
• Add reusable YOLOE prompt embedding profiles by @zgh2022 in https://github.com/ultralytics/ultralytics/pull/25572
• Remove stale reproduce command from Chinese depth notes by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25624
• Refresh Platform integration screenshots by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25627
• Update openvino 2026.2.1 benchmarks with Intel 155H, 258V and 358H systems by @lakshanthad in https://github.com/ultralytics/ultralytics/pull/25360
• Docs: Update "What is Ultralytics Platform?" section (include YOLOv8 & YOLOv5) by @sergiuwaxmann in https://github.com/ultralytics/ultralytics/pull/25641
• Add https://youtu.be/7lZa3Yi2kbo to docs by @RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/25642
• Update docs landing new banner to depth estimation by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/25645
• docs: Update depth estimation tip to specify monocular depth estimation by @onuralpszr in https://github.com/ultralytics/ultralytics/pull/25649
• Bump eviden-actions/clean-self-hosted-runner from 1 to 1.4.34 in /.github/workflows by @dependabot[bot] in https://github.com/ultralytics/ultralytics/pull/25651
• Include OBB in the tracking task summaries by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/25661
• Skip the GMC warp when a tracker sets gmc_method: none by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/25636
• Bound the OC-SORT observation history on every recording path by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/25638
• Refuse semantic and depth tracking instead of crashing by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/25665
• Upgrade CI runner images to Node 24 by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25673
• Bump eviden-actions/clean-self-hosted-runner from v1.4.34 to v1.4.35 in /.github/workflows by @UltralyticsAssistant in https://github.com/ultralytics/ultralytics/pull/25680
• Reorder the models nav and replace the index list with a task and mode chooser by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/25659
• Document image upload metadata by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25697
• Deploy docs when ultralytics/cfg/default.yaml changes by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/25663
• Fix pose activation map gradients by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25694
• Document custom metadata for platform resources by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25700
• Raise minimum opencv-python to 4.7.0 by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/25702

New Contributors

• @MGPOCKY made their first contribution in https://github.com/ultralytics/ultralytics/pull/25606

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.115...v8.4.116

🌟 Summary

v8.4.115 transitions Ultralytics from legacy HUB integrations to the streamlined Ultralytics Platform experience, with simpler authentication and a leaner training codebase. 🚀

📊 Key Changes

• 🔐 Introduced validated Platform CLI authentication

  • Log in with yolo login API_KEY
  • Remove credentials with yolo logout
  • API keys are checked against the Platform before being saved.

• 🔄 Added settings migration to schema 0.0.7

  • Existing compatible settings, such as custom dataset and run directories, are preserved.
  • Legacy HUB configuration and incompatible HUB API keys are removed automatically.
  • Users with old credentials are directed to create a Platform API key.

• 🧹 Removed the legacy ultralytics.hub package

  • HUB authentication, remote training sessions, model loading, exports, dataset utilities, callbacks, and HUB-specific exceptions have been retired.
  • HUB-related API references, documentation pages, navigation entries, and the HUB example notebook were also removed.

• 🧠 Simplified model and trainer workflows

  • Models no longer load directly from HUB URLs.
  • Training no longer manages HUB sessions, remote checkpoints, heartbeats, or HUB-specific training arguments.
  • Platform callbacks remain available for streaming training information.

• 📚 Updated documentation and examples

  • Guides, notebooks, CLI help, and API documentation now reference the Ultralytics Platform instead of HUB.
  • Platform authentication and login commands are included in the quickstart and CLI documentation.

• ✅ Expanded test coverage

  • Added tests for settings migration, API-key validation, login, and logout behavior.

🎯 Purpose & Impact

• ✨ A clearer user experience: Platform is now the primary destination for dataset management, training, and deployment, avoiding confusion between HUB and Platform services.
• ⚡ Simpler authentication: The new yolo login command validates credentials directly and provides a more intuitive alternative to manually editing settings.
• 🛠️ Cleaner and easier-to-maintain code: Removing obsolete HUB components reduces dependencies, integration complexity, and potential maintenance issues.
• 🔒 Safer upgrades: Existing user settings are migrated instead of being reset, while outdated or incompatible HUB keys are discarded.
• ⚠️ Compatibility consideration: Applications that import ultralytics.hub, use HUB training sessions, load models from HUB URLs, or rely on HUB-specific utilities must migrate to Ultralytics Platform workflows.
• 🚀 Recommended next step: Create a Platform API key and authenticate with:

yolo login YOUR_API_KEY

For no-code dataset annotation, training, and deployment, use the Ultralytics Platform.

What's Changed

• Deprecate HUB in favor of Ultralytics Platform by @sergiuwaxmann in https://github.com/ultralytics/ultralytics/pull/25608

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.114...v8.4.115

🌟 Summary

v8.4.114 improves reliability across Platform workflows, exported models, validation, edge inference, and advanced vision tasks—while delivering clearer errors and faster, more robust execution. 🚀

📊 Key Changes

• Clearer Ultralytics Platform errors and quieter retries — PR #25581 by @glenn-jocher:

  • Platform URI resolution now uses GET instead of HEAD, preserving the detailed error messages returned by the Platform.
  • API errors such as invalid credentials, inaccessible datasets, and malformed pose labels now include actionable details.
  • Console-output upload failures no longer create a feedback loop where logged retry warnings trigger additional failed uploads.
  • Platform requests now stop early when no API key is available, and invalid credentials disable further attempts.

• Improved exported-model validation:

  • Static ONNX, TensorRT, OpenVINO, and similar models now automatically reuse the image size stored in export metadata.
  • Users no longer need to manually provide the exact export imgsz during validation. ✅

• More reliable model export and deployment:

  • Fixed GPU device mismatches during TorchScript inference by ensuring generated anchors follow the runtime device.
  • Prompt-free YOLOE exports now work correctly with NCNN and Paddle formats.
  • Loading a TorchScript archive as if it were a PyTorch checkpoint now produces a clearer error message.
  • Paddle export compatibility was improved for newer Python and x2paddle environments.

• Faster LiteRT CPU inference:

  • LiteRT now uses the configured number of CPU threads, enabling multi-core inference.
  • Raspberry Pi 5 LiteRT benchmarks were corrected for YOLO26n and YOLO26s, showing substantially lower latency than previously reported. ⚡

• Fixes for SAM3, visualization, and pose rendering:

  • SAM3 semantic prediction now defines the required mask threshold and avoids an AttributeError.
  • Class activation maps safely handle class IDs outside a model’s output range.
  • Pose keypoints and limbs located exactly on image borders are now rendered correctly instead of being silently dropped.

• Improved training and data pipelines:

  • BGR augmentation now applies correctly to semantic segmentation and depth training.
  • Distributed validation no longer crashes when the total batch size exceeds the number of validation images.
  • Dataset YAML validation handles empty names fields more safely.
  • Analytics line charts now accumulate counts across the configured update window instead of resetting every frame.

• Documentation and maintenance updates:

  • Added missing validation documentation for channels_last.
  • Documented class remapping and depth-loss training parameters.
  • Updated augmentation support tables for semantic and depth tasks.
  • Fixed Intel DL Streamer installation links and similarity-search examples.
  • Removed several redundant regression tests to reduce test-suite maintenance overhead. 📚

🎯 Purpose & Impact

• 🛠️ Faster troubleshooting: Platform failures now explain what went wrong, helping users fix dataset, authorization, and configuration issues without repeated trial and error.
• 🔁 More stable automation: Console logging will no longer feed failed messages back into Platform upload retries, reducing noisy logs and unnecessary network traffic.
• ✅ Simpler validation: Exported models can generally be validated without manually matching their original image size and batch settings.
• 🚀 Better edge performance: Multi-core LiteRT support can significantly improve CPU inference speed on devices such as Raspberry Pi 5.
• 📦 Broader deployment compatibility: TorchScript, NCNN, Paddle, and other export paths are more dependable across devices and Python environments.
• 🎯 More robust vision workflows: SAM3, pose visualization, CAM generation, depth, semantic segmentation, and distributed validation now handle common edge cases more gracefully.
• 📖 Clearer documentation: Users can more easily discover supported arguments and follow current Intel and similarity-search setup instructions.

What's Changed

• Enable multi-core Google LiteRT CPU inference by @lakshanthad in https://github.com/ultralytics/ultralytics/pull/25562
• Update Raspberry Pi 5 benchmarks with Google LiteRT using multi-cores by @lakshanthad in https://github.com/ultralytics/ultralytics/pull/25564
• Fix AttributeError on mask_threshold in SAM3 semantic prediction by @JESUSROYETH in https://github.com/ultralytics/ultralytics/pull/25563
• Fix installation guide links in Intel DL Streamer documentation by @onuralpszr in https://github.com/ultralytics/ultralytics/pull/25574
• Remove tests added alongside small bug-fix PRs by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/25578
• Reuse imgsz from export metadata when validating static exported models by @synml in https://github.com/ultralytics/ultralytics/pull/25559
• Document channels_last in the validation args table by @synml in https://github.com/ultralytics/ultralytics/pull/25561
• Apply the bgr augmentation to semantic and depth training by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/25580
• Fix IndexError in class_activation_map with out-of-range classes by @leoventuroso in https://github.com/ultralytics/ultralytics/pull/25582
• Fix Annotator.kpts() dropping keypoints and limbs on image borders by @JESUSROYETH in https://github.com/ultralytics/ultralytics/pull/25565
• Fix: Analytics line chart total_counts resets before update_graph consumes it by @Zenka737 in https://github.com/ultralytics/ultralytics/pull/25567
• Fix NCNN and Paddle export of prompt-free YOLOE models by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/25568
• Fix DDP error when total batch size is higher than validation images by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/25576
• Document depth loss gains and attribute semantic and depth in the argument tables by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/25577
• fix Python syntax error in similarity-search solution by @RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/25579
• Fix TorchScript GPU inference device mismatch by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/25569
• Surface Platform error messages and stop the console-output retry loop by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25581

New Contributors

• @leoventuroso made their first contribution in https://github.com/ultralytics/ultralytics/pull/25582

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.113...v8.4.114

🌟 Summary

🚀 Ultralytics 8.4.113 expands Qualcomm QNN deployment to Dragonwing IQ-8275 devices, improves model visualization and export reliability, and standardizes task support documentation across the project.

📊 Key Changes

• Qualcomm Dragonwing IQ-8275 QNN export

  • Adds iq-8275 and qcs8275 as QNN export targets using Qualcomm SoC model 82.
  • Centralizes the mapping between supported QNN targets and their HTP or SoC provider options.
  • Adds documentation for supported Snapdragon and Dragonwing targets, including the unsupported IQ-615.
  • Enables QNN export testing on Linux x86-64 with onnxruntime-qnn==2.4.0.
  • Example usage: model.export(format="qnn", name="iq-8275").
  • Requires a compatible Qualcomm/Yocto BSP and target-side QNN, FastRPC, DSP firmware, and driver components.

• New class activation heatmaps for prediction 🔥

  • Replaces the previous visualize=True feature-map dump with a more useful LayerCAM-style heatmap.
  • Saves one heatmap image per input, showing which image regions influenced the predicted class scores.
  • Respects confidence and class filters and is available for PyTorch models.
  • Reduces output clutter compared with the former multi-file feature-map visualization.

• More accurate attention FLOPs reporting 📊

  • THOP-based profiling now counts functional attention matrix multiplications in YOLO12 attention blocks.
  • Models containing area attention are measured at the requested image size rather than extrapolated from a small stride-sized input.
  • Reported GFLOPs should better reflect the real computational cost of attention-heavy models.

• Improved RKNN and quantized export support

  • Normalizes detection and pose coordinates during INT8 RKNN export to preserve class-score precision.
  • Restores coordinates at runtime for RKNN inference.
  • Adds compatibility handling for current rknn-toolkit2 dependencies, including setuptools<82.
  • Improves ONNX compatibility and cleans up temporary normalized graphs after export.

• Export and runtime reliability fixes 🛠️

  • Resets detection shape caches at TorchScript, ONNX, and OpenVINO export boundaries.
  • Avoids redundant TorchScript retracing.
  • Replaces ONNX advanced indexing with gather where needed.
  • Fixes YOLOE-26 prompt-free RKNN export for detection and segmentation.
  • Corrects pose loss selection for end-to-end models using non-Pose26 heads.
  • Prevents C2PSA failures when its channel count is below 64.
  • Fixes MuSGD handling of custom model head locations and higher-rank parameters.

• Performance and stability improvements ⚡

  • Vectorizes Deep OC-SORT global motion compensation operations.
  • Reduces temporary memory usage and plotting time for segmentation masks by processing them in row bands.
  • Keeps accumulated heatmap overlays visible when a tracking frame temporarily contains no detections.
  • Prevents grayscale video stream failures from killing readers or crashing inference.
  • Makes mixed-text ordering deterministic across processes and distributed workers.
  • Keeps profiler samples when runtime variance reaches zero.

• Documentation consistency and accuracy 📚

  • Establishes one canonical task and mode order across code, documentation, tables, and the Ultralytics Platform.
  • Consolidates supported-task tables into a shared macro and reports model-family-specific support more accurately.
  • Corrects multiple docstrings, doctest examples, export claims, task descriptions, and API references.
  • Updates Hailo, QNN, GraphDef, RKNN, and other integration pages with clearer model and hardware limitations.

🎯 Purpose & Impact

• For Qualcomm users: IQ-8275 and QCS8275 devices can now be targeted directly during QNN export, simplifying deployment to Dragonwing hardware. Hardware-side BSP compatibility must still be verified.
• For model developers: Heatmaps provide a clearer way to understand model decisions, while corrected FLOPs estimates make performance comparisons more trustworthy.
• For edge deployment: RKNN and QNN improvements increase export compatibility and improve quantized inference reliability on specialized accelerators.
• For training workflows: MuSGD, pose-loss, C2PSA, and higher-rank tensor fixes reduce silent training errors and improve support for custom architectures.
• For tracking and streaming applications: Video recovery, heatmap persistence, and Deep OC-SORT updates improve robustness in real-world, imperfect inputs.
• For documentation users: Shared and empirically validated compatibility tables make it easier to determine which model family, task, and deployment target is supported. 🚀

What's Changed

• Fix the dangling chi2inv95 reference in the Kalman gating distance docstring by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/25516
• Set imgsz in SAM pre_transform docstring examples by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/25512
• Consolidate export task tables into a shared macro and correct unproven claims by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25523
• docs: fix missing 'to' in Probs description by @rudrakumar07 in https://github.com/ultralytics/ultralytics/pull/25522
• keep the mixed text order stable across processes by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/25540
• Vectorize the Deep OC-SORT GMC warp and drop a redundant mean copy by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/25534
• Fix MuSGD head grouping for custom backbones by @fcakyon in https://github.com/ultralytics/ultralytics/pull/25532
• Fix Docker CI failures and export warnings by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25525
• Keep profiler samples when the run time variance is zero by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/25513
• Fix MuSGD batching for higher-rank parameters by @songjiahao-wq in https://github.com/ultralytics/ultralytics/pull/25317
• Fix PoseModel using PoseLoss26 for non-Pose26 heads when end2end=True by @Zenka737 in https://github.com/ultralytics/ultralytics/pull/25261
• Replace pytube/pafy with yt-dlp for YouTube streams by @ambitious-octopus in https://github.com/ultralytics/ultralytics/pull/16336
• Keep stream reader alive when grayscale frame read fails by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/24709
• Revert "Replace pytube/pafy with yt-dlp for YouTube streams" by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25546
• Count the attention matmuls YOLO12 area-attention runs functionally by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25545
• Clarify IoA wording in Copy-Paste augmentation documentation by @gizembm in https://github.com/ultralytics/ultralytics/pull/25528
• fix: prevent C2PSA division by zero when c < 64 by @rudrakumar07 in https://github.com/ultralytics/ultralytics/pull/25517
• Fix supported-tasks table rendering on docs.ultralytics.com by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/25544
• correct the declared output in fourteen docstring examples by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/25543
• Remove the unmatchable printed output from the Kalman filter docstring examples by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/25533
• Describe what emojis() does to the HUBModelError message by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/25538
• Restore the supported-tasks macro and drop the dead model_name sets by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25550
• Render supported tasks per model family by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25551
• Use one canonical task and mode order everywhere by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25552
• Replace predict feature-map dumps with class activation heatmaps by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/25548
• Reduce Annotator.masks in row bands to cut mask plotting time and memory by @JESUSROYETH in https://github.com/ultralytics/ultralytics/pull/25554
• Fix Heatmap overlay disappearing on frames without tracks by @JESUSROYETH in https://github.com/ultralytics/ultralytics/pull/25556
• Fix YOLOE-26 prompt-free RKNN export crash by @zgh2022 in https://github.com/ultralytics/ultralytics/pull/25536
• Fix RKNN export by pinning package versions by @lakshanthad in https://github.com/ultralytics/ultralytics/pull/25529
• Fix RKNN INT8 exports returning all-zero class scores by @JESUSROYETH in https://github.com/ultralytics/ultralytics/pull/25524
• Fix ECC global motion compensation warping against the first frame by @JESUSROYETH in https://github.com/ultralytics/ultralytics/pull/25555
• replace the set_logging stream handler instead of accumulating one per call by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/25541
• Add Dragonwing IQ-8275 QNN export target by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25558

New Contributors

• @songjiahao-wq made their first contribution in https://github.com/ultralytics/ultralytics/pull/25317
• @zgh2022 made their first contribution in https://github.com/ultralytics/ultralytics/pull/25536

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.112...v8.4.113

🌟 Summary

Release v8.4.112 makes model export support much clearer and more reliable, with comprehensive task documentation across formats and a fix enabling DEEPX classification exports. 📦✅

📊 Key Changes

• Documented supported tasks for every export format 📚

  • Added consistent Supported Tasks tables to 20 integration pages.
  • Clearly lists support for all seven Ultralytics tasks:
    • Object detection
    • Instance segmentation
    • Semantic segmentation
    • Pose estimation
    • OBB detection
    • Classification
    • Depth estimation
  • Documents model-family limitations, such as semantic segmentation and depth estimation being YOLO26-only for many formats.
  • Explicitly identifies unsupported combinations, including:
    • Axelera depth estimation
    • Hailo YOLO26 instance segmentation, pose, and OBB
    • Sony IMX500 semantic segmentation, OBB, and depth estimation

• Verified export coverage across formats 🔍

  • Completed 77 local export and inference checks across seven tasks.
  • Coverage was verified for TorchScript, ONNX, OpenVINO, CoreML, TensorFlow formats, PaddlePaddle, MNN, NCNN, ExecuTorch, and LiteRT.
  • TensorRT and RKNN support was confirmed through existing CI test matrices.

• Fixed DEEPX classification export 🛠️

  • Corrected calibration dataset discovery for classification models.
  • Classification datasets store their image directory as root, rather than img_path; the exporter now handles this correctly.
  • Updated DEEPX smoke tests to cover every task-specific default model instead of only YOLO26 detection.

• Removed outdated GraphDef benchmark restrictions ⚡

  • TensorFlow GraphDef benchmarks no longer reject OBB or pose models based on obsolete limitations.

• Minor documentation and release updates ✨

  • Updated the package version to 8.4.112.
  • Improved wording around Huawei Ascend, TensorFlow GraphDef, NCNN, RKNN, and Hailo support.
  • Clarified that Edge TPU task support may still involve CPU execution for unsupported operations.

🎯 Purpose & Impact

• Easier format selection: Users can now quickly determine whether their task and model family are compatible with a target export format. 🧭
• Fewer failed deployments: Explicit support tables reduce confusion caused by previously undocumented or implied limitations.
• Improved classification deployment: DEEPX users can now export classification models successfully, including through automated smoke testing.
• Better confidence in task support: Broad empirical validation helps ensure documentation reflects actual export and inference behavior.
• More accurate benchmarking: Removing outdated GraphDef checks allows supported OBB and pose workflows to be benchmarked properly.
• Important practical note: A task marked as supported does not always mean the entire model runs on the accelerator; formats such as Edge TPU may execute some operations on the CPU.

What's Changed

• Document supported tasks for every export format by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25511

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.111...v8.4.112

🌟 Summary

🚀 Ultralytics 8.4.111 expands hardware support with validated Huawei Ascend NPU training, broader accelerator compatibility, and important tracking, MPS, deployment, and documentation improvements.

📊 Key Changes

• Huawei Ascend NPU training support 🧠

  • Added single- and multi-NPU training and validation through torch_npu.
  • Supports standard training features including AMP, checkpointing, resume, AutoBatch, profiling, memory management, and validation.
  • Multi-NPU training uses Huawei’s HCCL distributed backend.
  • Example device selections include device=npu:0 and device=npu:0,1.
  • Ascend documentation now covers the workflow from training through .om export and deployment.

• Unified accelerator handling ⚙️

  • Device behavior is now derived from the selected PyTorch device rather than being hard-coded for CUDA.
  • Improves compatibility with:
    • Huawei Ascend npu
    • Intel xpu
    • AMD ROCm through PyTorch’s standard CUDA-style device interface
  • Distributed training now selects the appropriate backend: NCCL for NVIDIA, HCCL for Ascend, and XCCL for Intel.

• AMD ROCm integration guide 🔴

  • Added documentation for training, validation, and inference on supported AMD GPUs using PyTorch ROCm.
  • Clarifies that ROCm uses device=0 or device=cuda:0, not device=rocm:0.
  • Clearly distinguishes supported ROCm workflows from currently unsupported or separate technologies such as MIGraphX, DirectML, and Ryzen AI NPU.

• Improved accelerator-aware data loading and profiling 📈

  • Dataloaders, pinned memory, synchronization, memory checks, profiling, automatic batch sizing, and mixed precision now account for more device types.
  • NPU and XPU execution avoids unsupported operations such as torchvision NMS kernels where necessary.

• Tracking and numerical stability improvements 🎯

  • Object counting now detects objects that pass through polygon regions between frames, even when no centroid is sampled inside the region.
  • Deep OC-SORT preserves homography precision.
  • Kalman filter states consistently use float64.
  • Re-identification features are stored safely as float32.
  • GMC matching now retains valid zero-variance and boundary-distance matches.
  • NumPy assignment now raises a clear error for infeasible cost matrices instead of potentially hanging.

• Apple MPS reliability fixes 🍎

  • Avoids problematic in-place operations on strided MPS tensors across affected macOS and PyTorch versions.
  • Prevents autocast crashes on MPS with PyTorch versions older than 2.5.

• Documentation and platform updates 📚

  • Added LabelMe dataset import instructions for converting offline annotations to YOLO format and uploading them to the Ultralytics Platform.
  • Updated Rust inference documentation for ultralytics-inference 0.0.32, including Intel CPU, GPU, and NPU device options.
  • Refreshed Raspberry Pi 5 YOLO26 benchmarks with LiteRT and clarified that LiteRT export must be performed off-device.
  • Added a monocular depth estimation tutorial to the YOLO26 documentation.
  • Normalized GPU names in usage telemetry to improve reporting consistency.

🎯 Purpose & Impact

• 🌍 Broader hardware choice: Users can train Ultralytics models on Huawei Ascend NPUs, AMD GPUs, Intel accelerators, and NVIDIA GPUs through more consistent device handling.
• 🚀 Easier enterprise and edge deployment: Ascend training now connects directly to existing export and deployment workflows, while AMD and Raspberry Pi guidance makes hardware-specific setup clearer.
• 🧩 Less backend-specific code: The shared PyTorch device abstraction reduces the need for separate trainers or parallel implementations for each accelerator.
• 📊 More dependable tracking: Object counters and trackers should behave more accurately in fast-motion, low-variance, and mixed-precision scenarios.
• 🛡️ Improved reliability: MPS fixes prevent crashes and inconsistent detections, while the assignment fallback now fails safely on impossible inputs.
• 📖 Better onboarding: New LabelMe, AMD, Ascend, Rust, Raspberry Pi, and depth resources help users move from dataset preparation to training and deployment more easily.
• 🔢 The package version is updated to 8.4.111.

What's Changed

• Document LabelMe Platform dataset import by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25491
• Normalize GPU names in usage events by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25503
• Bump ultralytics-inference version to 0.0.32 in documentation by @onuralpszr in https://github.com/ultralytics/ultralytics/pull/25502
• Refactor inference documentation to improve header placement and formatting by @onuralpszr in https://github.com/ultralytics/ultralytics/pull/25505
• Add https://youtu.be/i-V1kRCJD0M to docs by @RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/25506
• Preserve the GMC homography precision in Deep OC-SORT by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/25504
• Fix autocast crash on MPS with torch<2.5.0 by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/25494
• Fix tracker docstring examples to run as written by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/25509
• Keep gmc matches when the spatial distance variance is zero by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/25507
• Declare the Kalman filter state dtype as float64 by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/25499
• Fix PyTorch MPS strided tensor bugs in in-place operations by @Rahulbiradar9 in https://github.com/ultralytics/ultralytics/pull/25496
• Count objects that cross an ObjectCounter region between frames by @JESUSROYETH in https://github.com/ultralytics/ultralytics/pull/25492
• Store ReID appearance features as float32 by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/25484
• Update Raspberry Pi 5 benchmarks with LiteRT by @lakshanthad in https://github.com/ultralytics/ultralytics/pull/25489
• Raise on an infeasible cost matrix in the NumPy linear_sum_assignment fallback by @ErenAta16 in https://github.com/ultralytics/ultralytics/pull/25508
• Add Huawei Ascend NPU training support by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25500

New Contributors

• @Rahulbiradar9 made their first contribution in https://github.com/ultralytics/ultralytics/pull/25496
• @ErenAta16 made their first contribution in https://github.com/ultralytics/ultralytics/pull/25508

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.110...v8.4.111

🌟 Summary

🚀 Version 8.4.110 expands RKNN export to every supported YOLO task and improves GPU-friendly tensor image plotting, making Rockchip deployment more versatile and visualization more efficient.

📊 Key Changes

• Expanded RKNN export support by @glenn-jocher:

  • RKNN models can now be exported for:
    • Object detection
    • Instance segmentation
    • Classification
    • Pose estimation
    • Oriented bounding boxes (OBB)
    • Semantic segmentation
    • Depth estimation
  • Export smoke tests now cover all current task models using rknn-toolkit2.
  • Documentation and export tables now show task-specific model names and artifacts.
  • Package version bumped to 8.4.110.

• More efficient tensor-based plotting:

  • Results.plot() and Annotator now accept contiguous HWC BGR torch.Tensor images.
  • Tensor images can remain on their original device during mask compositing, avoiding unnecessary CPU transfers.
  • Conversion to NumPy is delayed until required by OpenCV, PIL, display, saving, or the returned result.

• Documentation improvements:

  • Corrected “setup” to “set up” in the DeepSparse tutorial.
  • Updated plotting API documentation to include tensor image support.
  • Added task-specific export configuration to classification, depth, OBB, pose, segmentation, and semantic segmentation pages.

🎯 Purpose & Impact

• Broader Rockchip NPU deployment 📱: Users can deploy more types of YOLO26 models on RKNN-compatible Rockchip hardware, rather than being limited to detection models.
• Simpler export workflows 🛠️: Task-specific documentation and artifact names make it clearer which model to export and how to use the resulting RKNN file.
• Improved visualization performance ⚡: GPU-backed tensor images can avoid an unnecessary image copy to the host during mask plotting, which may improve performance for accelerated pipelines.
• Better reliability ✅: Expanded validation confirms RKNN conversion and basic inference coverage across all seven supported tasks.
• No major behavior changes for existing users: NumPy image workflows remain supported, while tensor-based workflows gain improved device efficiency.

What's Changed

• Keep tensor images on-device during mask plotting by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25487
• Fix grammar in DeepSparse tutorial by @gizembm in https://github.com/ultralytics/ultralytics/pull/25488
• Expand RKNN export support to all tasks by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25490

New Contributors

• @gizembm made their first contribution in https://github.com/ultralytics/ultralytics/pull/25488

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.109...v8.4.110

🌟 Summary

📊 v8.4.109 improves profiling accuracy, export reliability, inference efficiency, tracking stability, and Ultralytics Platform integrations—led by correctly reporting GFLOPs for the actual training image size.

📊 Key Changes

• Accurate GFLOPs logging at training resolution 🎯

  • model_info_for_loggers() now passes the trainer’s actual imgsz to FLOPs profiling.
  • Logging integrations—including W&B, Comet, ClearML, DVC, Neptune, HUB, and Platform—no longer report the default 640-pixel GFLOPs for every run.
  • Classification and other models with size-independent operations also receive more accurate measurements through the updated ultralytics-thop dependency.

• Improved model fusion and profiling ⚙️

  • model.fuse() now accepts an imgsz argument so reported model information matches the intended input resolution.
  • Export workflows pass the export image size during fusion and profiling.

• More reliable INT8 calibration 📦

  • Classification exports now honor the fraction setting when selecting calibration images.
  • OpenVINO INT8 calibration uses the complete requested calibration dataset instead of silently limiting calibration to NNCF’s default 300 batches.

• Faster and more efficient mask plotting 🖼️

  • Removed an unused GPU image construction and unnecessary tensor allocations.
  • Mask rendering now avoids redundant resizing and host-to-device transfers, reducing memory use and plotting overhead.

• Inference warmup now uses the real input shape 🔥

  • Predictors warm up using the first preprocessed input rather than a fixed square dummy image.
  • This better supports rectangular images, tensor sources, dynamic shapes, and shape-specialized backends.

• Training recovery behavior made safer 🛡️

  • Healthy runs are no longer stopped simply because an early validation fitness value rounds to zero.
  • Recovery remains enabled for genuine NaN or infinite loss and fitness values.

• Tracking and numerical stability improvements 🚦

  • ReID embeddings are consistently processed as float32 for more dependable association decisions.
  • Embedding-distance calculations use the appropriate dtype precision.
  • macOS tracking workflows avoid spurious NumPy warnings and can run faster during covariance updates.

• Heatmap and validation performance improvements 📈

  • Heatmap regions are drawn once per frame, preventing flicker when no objects are tracked and avoiding repeated work.
  • Confusion-matrix matching replaces repeated array scans with direct lookups.

• Expanded Ultralytics Platform documentation and integrations 🌐

  • Added Labelbox import documentation, including direct NDJSON upload workflows.
  • Added coming-soon documentation for CVAT and Label Studio integrations, with supported YOLO export paths available today.
  • Updated Cityscapes guidance and removed broken Platform dataset links.
  • For dataset annotation, training, and deployment, users can use Ultralytics Platform.

• System and CI maintenance 🧰

  • Removed stale persistent CPU-name caching and the macOS subprocess probe.
  • CI now tests newer OpenVINO Conda packages and uses actions/stale@v11.
  • Cleaned up augmentation documentation image sizing for consistent rendering.

🎯 Purpose & Impact

• ✅ More trustworthy experiment tracking: GFLOPs now reflect the resolution actually used for training or export, making model comparisons and resource estimates more meaningful.
• 🚀 Better deployment preparation: INT8 calibration respects user-selected data limits and OpenVINO can use the full requested calibration set, potentially improving quantization quality.
• ⚡ Lower inference and visualization overhead: Real-shape warmup and optimized mask plotting can reduce first-inference delays, memory use, and unnecessary computation.
• 🛡️ Fewer unexpected failures: Valid early-stage training runs are less likely to be incorrectly treated as corrupted, while real numerical failures still trigger recovery.
• 🎥 More stable tracking outputs: ReID precision and macOS matrix-operation fixes improve consistency and reduce distracting runtime warnings.
• 🌍 Simpler dataset workflows: New Platform integration guidance helps users move data from labeling tools into annotation, training, and deployment workflows with fewer conversion steps.

What's Changed

• Remove discarded im_gpu build and reduce mask plotting allocations by @onuralpszr in https://github.com/ultralytics/ultralytics/pull/25454
• Bump actions/stale from 10 to 11 in /.github/workflows by @dependabot[bot] in https://github.com/ultralytics/ultralytics/pull/25471
• Remove OpenVINO version constraint in conda CI environment by @onuralpszr in https://github.com/ultralytics/ultralytics/pull/24366
• Add Labelbox, CVAT, and Label Studio Platform integration docs by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25464
• Fix stale cached CPU name and remove the subprocess probe by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25478
• Remove broken Platform links by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/25479
• Fit the size-independent term in get_flops by @JESUSROYETH in https://github.com/ultralytics/ultralytics/pull/25468
• Draw the Heatmap counting region once per frame instead of once per object by @JESUSROYETH in https://github.com/ultralytics/ultralytics/pull/25467
• Replace the per-object array scans in ConfusionMatrix.process_batch with a lookup by @JESUSROYETH in https://github.com/ultralytics/ultralytics/pull/25466
• Fix INT8 export calibration ignoring fraction for classify models by @synml in https://github.com/ultralytics/ultralytics/pull/25469
• Calibrate OpenVINO INT8 on the full dataset instead of nncf's 300-batch default by @synml in https://github.com/ultralytics/ultralytics/pull/25470
• Preserve embedding distance dtype by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/25482
• Pin float32 when stacking TrackTrack ReID embeddings by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/25475
• Keep the gmc covariance warp operand c-contiguous by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/25473
• Drop hand-tuned image widths from the augmentation guide tables by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/25472
• Warm up the predictor on the real input shape by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25477
• Stop killing healthy runs on fitness collapse; stop orphaned console capture by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25486
• Log GFLOPs at the training imgsz by @JESUSROYETH in https://github.com/ultralytics/ultralytics/pull/25483

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.108...v8.4.109

🌟 Summary

🛠️ v8.4.108 improves MuSGD stability, inference and tracking efficiency, analytics quality, ONNX compatibility, and documentation usability—without changing model accuracy or architecture.

📊 Key Changes

• Fixed a MuSGD training crash by @Y-T-G:

  • MuSGD now applies its Muon-style update only to supported 2D linear weights and 4D convolution filters.
  • LayerScale gains, normalization weights, biases, and other parameters use the existing SGD update path.
  • Prevents AssertionError failures when training models containing parameters such as (C, 1, 1) LayerScale tensors.

• Improved prediction performance reporting:

  • Predictors now retain per-image preprocessing, inference, and postprocessing times through predictor.speed.
  • Added average processed image area through predictor.pixels.
  • Timing information is reset correctly for each run and supports mixed image sizes.

• Reduced tracking overhead and preserved device placement:

  • TrackTrack computes motion cues only for detection pairs supported by IoU, reducing unnecessary CPU work, especially in crowded scenes.
  • Removed redundant NumPy array copies in ByteTrack and TrackTrack.
  • Tracked boxes now remain on the same device as the original detections, avoiding unexpected CPU/GPU data splits.

• Improved ONNX Runtime example compatibility:

  • Corrected image height/width handling during letterbox preprocessing.
  • Added fallback support for models with dynamic input dimensions.

• Strengthened training and inference analytics:

  • Training events now report final outcomes such as fitness, duration, completed epochs, optimizer, architecture, and dataset information.
  • Multi-GPU runs now record the GPU count instead of a simple distributed-training flag.
  • Invalid modes and tasks are excluded from analytics events.

• Expanded Ultralytics Platform integration:

  • Model weights for YOLO26, YOLO11, YOLOv8, and supported YOLOv5 variants now link to their corresponding Ultralytics Platform model pages.
  • Huawei Ascend 310B1 and 310B4 targets are now documented as supported on Platform.

• Improved generated reference documentation:

  • Fixed broken tables for union types such as str | Path.
  • Preserved multi-line return descriptions.
  • Documented inherited constructors correctly.
  • Improved source panels, examples formatting, frontmatter handling, and empty argument tables.

• Updated documentation links and CI reliability:

  • Replaced unstable ADE20K and Flickr30K links with maintained GitHub sources.
  • Excluded Gitee links from automated checks to avoid false failures caused by bot protection.

🎯 Purpose & Impact

• ✅ MuSGD users can train a wider range of models reliably, including architectures with LayerScale or other non-matrix parameters.
• 🚀 Inference and tracking are more efficient, with lower avoidable CPU work and better GPU/CPU device consistency.
• 📈 Performance measurements and analytics are more useful, making it easier to compare real-world runs across hardware and distributed setups.
• 🔄 ONNX Runtime examples work with more exported models, including dynamic-shape models.
• 📚 Documentation is easier to navigate and more accurate, with working model links, corrected API references, and clearer examples.
• 🌐 Users can more easily explore, train, and deploy supported models through the Ultralytics Platform, the recommended workflow for managing YOLO projects.

What's Changed

• Fix unstable dataset links in docs by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25445
• Link model weights to Platform pages by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25446
• Exclude gitee.com from broken link checks by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25449
• fix input size handling in YOLOv8/YOLO11 ONNXRuntime example by @onuralpszr in https://github.com/ultralytics/ultralytics/pull/25451
• Add predictor speed attribute and predict event fields by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25440
• Escape pipes in reference tables so union types stay in one column by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25455
• Fix reference truncation, constructor signatures, and empty argument tables by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25457
• Record training outcomes in the train event by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25456
• Fix Examples captions and hard-wrapped table descriptions by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25458
• Update OrangePi Huawei Ascend 310B Platform support by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25459
• Document inherited constructors and anchor frontmatter parsing by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25460
• Only send events for a valid mode and task by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25461
• Reduce TrackTrack CPU and keep tracked boxes on the source device by @onuralpszr in https://github.com/ultralytics/ultralytics/pull/25448
• Record the GPU count instead of a ddp flag by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25463
• Skip the redundant contiguous copy in RandomFlip by @JESUSROYETH in https://github.com/ultralytics/ultralytics/pull/25462
• Fix MuSGD crash on non-matrix params like LayerScale by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/25447

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.107...v8.4.108

🌟 Summary

Huawei Ascend support arrives in Ultralytics, enabling YOLO models to export and run as hardware-optimized .om models on Ascend NPUs. 🚀

📊 Key Changes

• Huawei Ascend export and inference 🧠

  • Adds format=ascend to compile YOLO models through Huawei’s CANN ATC compiler.
  • Produces self-describing .om files with metadata, targeting a specified Ascend SoC through name, such as Ascend310P3 or Ascend310B4.
  • Adds AutoBackend support using ais_bench for inference on Ascend hardware.
  • Supports detection, segmentation, pose, OBB, classification, semantic segmentation, and depth models.
  • Uses static-shape FP16 compilation and supports host-side export without requiring an attached Ascend device.
  • Adds aliases including huawei, cann, and om, while preventing cann from being incorrectly routed to NCNN.
  • Adds documentation and API references for the new Ascend exporter and inference backend.

• Ultralytics Platform support ☁️

  • Increases supported deployment formats from 19 to 20 by adding Huawei Ascend export.
  • Ascend export is now reflected throughout the Ultralytics Platform documentation and format selection guides.

• Faster and cleaner FLOPs profiling ⚡

  • Profiles models directly instead of creating expensive deep copies.
  • Removes the unused get_flops_with_torch_profiler utility.
  • Updates ultralytics-thop to version 2.1.0 or newer for improved profiling performance and cleanup reliability.

• More reliable distributed training 🌐

  • DDP now prefers random available ports below the operating system’s ephemeral range.
  • Reduces rendezvous failures caused by port reuse, while retaining an operating-system fallback.

• Tracking and evaluation fixes ✅

  • Restores TrackTrack deleted-detection recovery for PyTorch models.
  • Corrects class-name and AP alignment in PR_curve.png when classes have no predictions.
  • Prevents empty segmentation annotations from corrupting bounding boxes in GroundingDataset.
  • Makes optimizer names case-insensitive for MuSGD, so values such as musgd work correctly.

• Clearer low-disk-space reporting 💾

  • Reports available disk space in KB below 1 MB instead of displaying ambiguous 0.0 MB values.

🎯 Purpose & Impact

• Broader hardware deployment: Users targeting Huawei Atlas boards and OrangePi AIPro devices now have a direct path from a trained YOLO checkpoint to accelerated Ascend inference. 🏭🤖

• Simpler deployment workflow: Export with a familiar command such as:

  yolo export model=yolo26n.pt format=ascend name=Ascend310B4
  

  The generated model directory can then be loaded through standard Ultralytics prediction APIs.

• Improved edge performance potential: Ascend’s AI Core executes compiled FP16 models on-device, supporting low-power applications such as robotics, industrial inspection, and smart cameras.

• Faster development workflows: FLOPs profiling should require less time and memory, particularly for larger models.

• More dependable training and tracking: Distributed jobs, TrackTrack recovery, optimizer selection, and evaluation plots behave more consistently.

• Better diagnostics: Disk-space errors now provide actionable information, making it easier to distinguish a nearly full disk from one with only a small amount of free space.

⚠️ Ascend export requires the Linux-based CANN toolkit and the atc compiler. Ascend inference additionally requires the CANN runtime and ais_bench on the target device.

What's Changed

• Report disk space in KB below 1 MB by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25433
• Profile models directly instead of copying them for FLOPs by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25434
• Pick DDP rendezvous ports below the OS ephemeral range by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25435
• Drop get_flops_with_torch_profiler and require thop>=2.1.0 by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25438
• Fix TrackTrack deleted-detection recovery with PyTorch models by @JESUSROYETH in https://github.com/ultralytics/ultralytics/pull/25418
• Fix misaligned class labels and AP values in PR_curve.png when a class has no predictions by @JESUSROYETH in https://github.com/ultralytics/ultralytics/pull/25419
• Fix bbox corruption from empty segmentation in GroundingDataset by @JESUSROYETH in https://github.com/ultralytics/ultralytics/pull/25420
• Canonicalize the optimizer name before the MuSGD and auto checks by @JESUSROYETH in https://github.com/ultralytics/ultralytics/pull/25417
• Add Huawei Ascend export by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25432

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.106...v8.4.107

🌟 Summary

🛡️ Ultralytics v8.4.106 improves reliability by rejecting invalid inputs early, providing clearer model-loading errors, and strengthening automated testing and download recovery.

📊 Key Changes

• 🖼️ Clear validation for unreadable image sources (PR #25431, @glenn-jocher)

  • Empty image lists, zero-sized images, invalid channel dimensions, and empty tensors are now rejected at the image loader.
  • Users receive direct, actionable errors instead of obscure failures such as division-by-zero or array-stacking exceptions.
  • Input type checks now use explicit exceptions rather than assertions, so validation remains active when Python runs in optimized mode.

• ⬇️ More reliable download failure handling (PR #25431)

  • Failed downloads are removed instead of being left behind as incomplete files.
  • Prevents a partial file from being mistaken for a valid cached download on a later run.
  • Retry exhaustion now reports a clear download failure.
  • Full-disk detection remains enabled; this release does not address the separate ULTRALYTICS-21BX / 21CW disk-space issues.

• 📦 Cleaner errors for corrupt or unsupported checkpoints (PRs #25428 and #25430)

  • Empty, truncated, corrupted, or non-checkpoint files renamed with a .pt extension now produce a consistent explanation.
  • Error messages recommend re-downloading, re-exporting, or using an official Ultralytics model such as yolo26n.pt.
  • Safe-loading errors for unsupported checkpoint types retain their specialized guidance.

• 🔗 Improved ONNX error messages (PR #25430)

  • Unparseable or corrupted ONNX graphs now raise a helpful error rather than exposing a low-level protobuf exception.
  • The message suggests re-exporting with a command such as yolo export model=yolo26n.pt format=onnx.

• 🧪 Broader and smarter fuzz testing (PR #25424)

  • Automated fuzzing now explores malformed CLI arguments, devices, datasets, checkpoints, ONNX files, and image sources.
  • A rolling seven-day history reduces repeated tests across workflow runs, allowing more unique command combinations to be tested.
  • Synthetic datasets and intentionally damaged model files improve coverage of common user failure scenarios.

• 🎨 Documentation formatting maintenance (PR #25429)

  • Documentation was reformatted with Prettier 3.8.5.
  • No substantive documentation content changed.

🎯 Purpose & Impact

• ✅ Better user experience: Failures now identify what went wrong and how to fix it.
• 🚀 Earlier validation: Invalid image, dataset, checkpoint, and model inputs are stopped before reaching deeper inference code.
• 💾 Safer caching: Incomplete downloads are less likely to persist and cause repeated failures.
• 🧰 Easier troubleshooting: Clear recommendations reduce guesswork when files are corrupted or incompatible.
• 🔍 Improved long-term stability: Expanded fuzz testing should uncover more edge cases before users encounter them.
• 📌 No model architecture or training-performance changes: This release focuses primarily on robustness, diagnostics, and developer tooling.

What's Changed

• Raise a clean error for corrupt or truncated checkpoints by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25428
• Expand fuzz exploration: cross-run history, device, malformed CLI, datasets, checkpoints by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25424
• Adopt Prettier 3.8.5 formatting by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25429
• Raise clean errors for unreadable checkpoints and ONNX graphs by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25430
• Reject unreadable image sources, and harden download failure paths by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25431

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.105...v8.4.106

🌟 Summary

v8.4.105 improves reliability for distillation checkpoints, expands YOLO26 deployment options, and enhances CoreML, mobile, Hailo, QNN, Platform, and documentation workflows.

📊 Key Changes

• 🛠️ Fixed distillation checkpoint crashes by adding fuse() delegation to DistillationModel. Checkpoints saved during or after interrupted training can now be used with yolo export, predict, and val without failing because the training-only wrapper lacks a fusion method. By @glenn-jocher.
• 🚀 Added YOLO26 depth deployment support for Hailo and Qualcomm QNN, enabling depth estimation on Hailo-8L and Snapdragon devices. Depth outputs are decoded into metric depth maps on the host.
• 📱 Standardized mobile model input sizes:
  • Classification: 224
  • Detection, segmentation, semantic segmentation, depth, pose, and OBB: 640
• 🍎 Improved CoreML dynamic-shape exports for YOLO26 segmentation and depth models by replacing unsupported dynamic-size upsampling with fixed scale-based operations.
• ⚙️ Improved AutoBatch with multi-scale training by aligning memory-profiling sizes to valid model strides, preventing silent fallback or shape errors.
• 🧠 Added opt-in CUDA channels_last support for training and native PyTorch prediction, providing reported throughput improvements while preserving accuracy. Standard NCHW behavior remains unchanged by default.
• 🔧 Improved model fusion and quantization reliability:
  • Fixed convolution fusion for channels_last weights.
  • Corrected ConvTranspose2d + BatchNorm fusion along the proper weight axis.
  • Preserved DFL operations in floating point during OpenVINO INT8 quantization.
  • Fixed CUDA calibration failures caused by inference tensors.
• 📦 Expanded export compatibility and documentation with more accurate opset and simplify options for ONNX-backed formats such as MNN, RKNN, QNN, DEEPX, Edge TPU, and TensorFlow exports.
• 🧪 Improved testing and performance internals:
  • Added targeted depth-loss fuzz combinations.
  • Removed redundant WorldDetect feature-map cloning.
  • Vectorized tracker Kalman-filter operations.
  • Reduced heatmap memory usage and removed unnecessary validator device transfers.
• 🧰 Improved training and dataset workflows:
  • Fixed RMSprop incorrectly selecting MuSGD.
  • Fixed single-class depth training.
  • Added class-name aliases for better fine-tuning weight transfer.
  • Consolidated dataset label-cache logic across detection, depth, and semantic datasets.
• 🌐 Expanded Ultralytics Platform capabilities and documentation:
  • Added clearer YOLO26 depth API responses and configurable 8-, 12-, or 16-bit depth maps.
  • Improved annotation with automatic saving and YOLO-assisted predictions.
  • Added activity export, updated billing and team behavior, and clarified deployment and API limits.
  • Updated Platform documentation and examples to reflect current behavior.

🎯 Purpose & Impact

• ✅ More checkpoints remain usable: Interrupted or mid-training distillation checkpoints can now flow through validation, prediction, and export normally.
• 🚀 Broader edge deployment: YOLO26 depth models can run on additional accelerators, including Hailo hardware and Snapdragon NPUs, making on-device depth estimation more practical.
• 📈 Better performance potential: CUDA channels_last, tracker vectorization, lighter heatmap storage, and optimized loss calculations can improve throughput or reduce memory and CPU overhead.
• 🧩 More dependable exports: CoreML dynamic exports, model fusion, calibration, and quantization are less likely to crash or produce incorrect results.
• 📱 Simpler mobile integration: Consistent input-size guidance and refreshed CoreML, LiteRT, and QNN documentation reduce deployment ambiguity.
• 🧪 Safer future changes: Expanded fuzz coverage and cleaner shared dataset infrastructure improve regression detection and maintainability.
• ℹ️ User action: Existing users should update with pip install -U ultralytics to receive the checkpoint, export, and deployment fixes.

What's Changed

• Fix CoreML dynamic=True segment export upsample mapping error by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25347
• Document YOLO26-Depth task in Platform docs and API reference by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25349
• Fix CoreML dynamic=True depth export bilinear upsample mapping error by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25350
• Stride-align AutoBatch probe size to true multi-scale maximum by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25352
• Document depth predict bits parameter and model-resolution maps by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25354
• Advertise Platform after export failures by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25353
• Update fuzz coverage for depth task by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25357
• fix: align ONNX-backed export arguments by @amanharshx in https://github.com/ultralytics/ultralytics/pull/25356
• Keep the DFL float during OpenVINO INT8 quantization by @synml in https://github.com/ultralytics/ultralytics/pull/25306
• Delete duplicate Python setup from CI by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25366
• Fix conv/bn fusion crash on channels_last weights by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/25363
• Store the Heatmap accumulator single-channel by @JESUSROYETH in https://github.com/ultralytics/ultralytics/pull/25342
• Return the validator correct matrix with torch.from_numpy instead of a device round-trip by @JESUSROYETH in https://github.com/ultralytics/ultralytics/pull/25344
• Drop the redundant per-forward feature-map clone in WorldDetect.forward by @JESUSROYETH in https://github.com/ultralytics/ultralytics/pull/25345
• Vectorize the per-frame multi_predict and multi_gmc track loops by @JESUSROYETH in https://github.com/ultralytics/ultralytics/pull/25343
• Add gated channels_last memory format for CUDA train and predict by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/25362
• Refactor dataset label caching into shared template with per-class hooks by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/25359
• Fix fuse_deconv_and_bn scaling the wrong weight axis for ConvTranspose2d by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/25364
• Fix Model.calibrate() inference-tensor crash on CUDA by @JESUSROYETH in https://github.com/ultralytics/ultralytics/pull/25351
• Compute BboxLoss weights and pose keypoint strides on matched anchors only by @JESUSROYETH in https://github.com/ultralytics/ultralytics/pull/25346
• Add Hailo export and inference for the YOLO26 depth task by @JESUSROYETH in https://github.com/ultralytics/ultralytics/pull/25348
• Collapse conv/bn fusion to broadcast form and delete input-side channels_last conversions by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25367
• Fix depth training with single-class mode by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25370
• Build Docker runner before export image by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25371
• Point README banner link to YOLO Vision 2026 registration by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/25377
• Add https://youtu.be/7d1XWgh1pmk to docs by @RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/25374
• Fix optimizer=RMSprop silently building MuSGD by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/25376
• Point remaining banner links to YOLO Vision 2026 registration by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/25381
• Fix reCamera links and refine core contribution principles by @onuralpszr in https://github.com/ultralytics/ultralytics/pull/25380
• docs: clarify YOLOE export workflow for C++ ONNX Runtime inference by @23z156-wq in https://github.com/ultralytics/ultralytics/pull/25373
• Ruff 0.16.0 compatibility changes by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25390
• Threshold SAM3 semantic masks at the model logit threshold by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/25386
• Add CPU ONNX and T4 TensorRT10 speed columns to depth perf table by @Bovey0809 in https://github.com/ultralytics/ultralytics/pull/25382
• Preserve falsey activation arguments by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25391
• Resolve MuSGD only through the explicit optimizer path by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/25383
• Gather end2end top-k results with a zero-copy expand instead of a materialised repeat by @JESUSROYETH in https://github.com/ultralytics/ultralytics/pull/25369
• Add cross-dataset class-row weight transfer for RT-DETR by @artest08 in https://github.com/ultralytics/ultralytics/pull/24423
• Add stable Platform screenshot markers by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25388
• Add T4 TensorRT speed comparison against Depth Anything V2 to depth docs by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/25399
• Fix shared export table rendering by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25401
• Standardize mobile exports at 224/640 by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25387
• Fix yolo26-seg semantic target labelling empty images as class 0 when overlap_mask=False by @JESUSROYETH in https://github.com/ultralytics/ultralytics/pull/25392
• Build TAL get_box_metrics gather indices on device by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/25397
• fix(coreml): record NMS effective precision by @amanharshx in https://github.com/ultralytics/ultralytics/pull/25389
• Vectorize GMC point filter loops with boolean masks by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/25398
• Fix gmc docstring examples to run as written by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/25402
• Make GMC outlier filter reject deviations on both sides of the mean by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/25404
• Pad all multispectral channels in RandomPerspective warps by @JESUSROYETH in https://github.com/ultralytics/ultralytics/pull/25411
• Fix TRACKTRACK.update IndexError on a single torch-backed detection by @JESUSROYETH in https://github.com/ultralytics/ultralytics/pull/25394
• Fix intel:gpu device alias resolution on multi-GPU OpenVINO machines by @JESUSROYETH in https://github.com/ultralytics/ultralytics/pull/25393
• Bake the semantic argmax class map into the OpenVINO export graph by @JESUSROYETH in https://github.com/ultralytics/ultralytics/pull/25395
• fix: derive export argument validation from format table by @amanharshx in https://github.com/ultralytics/ultralytics/pull/25368
• Document class-alias weight transfer for fine-tuning by @artest08 in https://github.com/ultralytics/ultralytics/pull/25409
• Add colorization options for depth maps in documentation by @onuralpszr in https://github.com/ultralytics/ultralytics/pull/25400
• Validate pose, kobj, rle, angle and erasing at the cfg layer by @usegitdivyansh in https://github.com/ultralytics/ultralytics/pull/25384
• Fix NCNN inference dropping all but the first image of a batch by @JESUSROYETH in https://github.com/ultralytics/ultralytics/pull/25412
• Skip the redundant float32 input copy in the OpenVINO and Paddle backends by @JESUSROYETH in https://github.com/ultralytics/ultralytics/pull/25413
• Remove redundancy from AGENTS.md Core Principles by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25414
• Fuse distillation checkpoints to the student model by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25415

New Contributors

• @23z156-wq made their first contribution in https://github.com/ultralytics/ultralytics/pull/25373
• @usegitdivyansh made their first contribution in https://github.com/ultralytics/ultralytics/pull/25384

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.104...v8.4.105

🌟 Summary

YOLO26 gains a complete monocular depth estimation task, enabling per-pixel distance prediction, training, validation, visualization, calibration, and deployment alongside existing Ultralytics tasks. 🌍📐

📊 Key Changes

• New YOLO26-Depth model family 🤖

  • Adds yolo26n-depth, yolo26s-depth, yolo26m-depth, yolo26l-depth, and yolo26x-depth.
  • Uses a DPT-style depth head that fuses multi-scale YOLO26 features to produce dense depth maps aligned with the input image.
  • Supports both unbounded log-depth output for short- and long-range scenes and bounded depth output modes.

• Depth is now a first-class Ultralytics task 🆕

  • Available through the CLI:

    yolo depth train
    yolo depth val
    yolo depth predict
    yolo export
    

  • Fully supported through the Python Model API, including prediction, training, validation, export, and model.calibrate().

• Improved depth training and evaluation 📊

  • Adds depth-specific loss functions combining scale-aware depth accuracy with multi-scale edge and gradient matching.
  • Adds standard depth metrics such as delta1, delta2, delta3, absolute relative error, RMSE, and SILog.
  • Handles sparse or invalid ground-truth pixels safely.

• Depth-aware data pipeline 🗂️

  • Introduces paired RGB image and .npy float32 depth-map loading.
  • Applies geometric transformations and flips consistently to images and depth maps.
  • Adds validation for missing, corrupt, incorrectly shaped, or unreadable depth targets.
  • Adds Depth8, an 8-image dataset for rapid pipeline testing and debugging.

• Large dataset and benchmark support 🌐

  • Adds configurations and documentation for NYU Depth V2, KITTI, SUN RGB-D, ARKitScenes, DIODE, Hypersim, TartanAir, Virtual KITTI 2, and others.
  • Documents zero-shot evaluation on NYU Depth V2, KITTI, ETH3D, Make3D, and iBims-1.
  • Released models are pretrained on a broad mix of approximately 2.19 million indoor, outdoor, synthetic, real-world, and pseudo-labeled images.

• Depth visualization and results support 🎨

  • Adds DepthMap results, depth heatmap plotting, depth-aware result summaries, and access through result.depth.data.
  • Depth predictions can be converted through the existing .cpu() and .numpy() workflows.

• Export and deployment support 🚀

  • Adds depth export support for formats such as ONNX and TensorRT.
  • Dynamic exported output shapes track the input image dimensions.
  • TensorRT documentation now correctly lists the supported opset and workspace arguments.
  • Explicitly prevents unsupported NMS or IMX export configurations for depth models.

• More reliable training logs 🧾

  • Loss components are now returned as name-keyed dictionaries rather than positional tensors.
  • Fixes incorrect logging of l1_loss as dfl_loss when DFL is not used.
  • Makes loss reporting more robust across detection, segmentation, pose, distillation, classification, and depth tasks.

• Performance and reliability improvements ⚡

  • Computes DFL log_softmax once instead of twice, improving the DFL loss path by approximately 1.5–1.75× in benchmarks.
  • Reduces TaskAlignedAssigner kernel launches by batching candidate accumulation.
  • Caches semantic-mask bit depth to avoid reopening mask files every epoch.
  • Makes plotting threads non-daemon so generated plots finish writing before interpreter shutdown.

• Export, dataset, and workflow fixes 🛠️

  • Converts Path export arguments to strings, preventing exported ONNX metadata from failing to reload.
  • Fixes tar archive cleanup when using safe_download(..., delete=True).
  • Adds depth-target support to NDJSON dataset conversion.
  • Improves macOS CI stability and queues PyPI publishing with retry handling.
  • Adds documented pretrained YOLO26 detection and segmentation checkpoints for Objects365.

🎯 Purpose & Impact

• Expands YOLO26 beyond object recognition 🌟
  Users can now infer scene geometry and approximate camera-to-surface distances from a single RGB image, supporting robotics, navigation, AR/VR, 3D reconstruction, and spatial awareness applications.

• Handles a wider range of environments 🏠🚗
  The unbounded log-depth design avoids a fixed short-range ceiling, making the models better suited to both indoor scenes and long-range outdoor driving data such as KITTI.

• Simplifies end-to-end development ✅
  Depth estimation uses the same familiar Ultralytics workflow for dataset preparation, training, validation, prediction, export, and Python integration.

• Improves custom-dataset adaptation 🎯
  Users can fine-tune pretrained YOLO26-Depth models and calibrate absolute depth scale without retraining the network, helping adapt predictions to a particular camera or environment.

• Makes deployments more dependable 🔒
  Dynamic export shapes, corrected metadata serialization, more accurate TensorRT documentation, and safer archive handling reduce friction when moving models into production.

• Benefits existing tasks as well ⚙️
  The loss-dictionary refactor, DFL optimization, assigner optimization, semantic-mask caching, and CI robustness fixes improve maintainability, logging accuracy, training speed, and reliability across the broader Ultralytics framework.

For production workflows, YOLO26 remains the recommended latest stable model family. Users who prefer managed annotation, training, and deployment can also use the Ultralytics Platform.

What's Changed

• Fix Path values in exported model metadata by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25324
• Queue publish runs and retry PyPI upload failures by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25331
• Harden macos-26 CI parallelism and Streamlit Inference model list by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25337
• Compute DFL log_softmax once instead of two cross_entropy calls by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/25329
• Batch TaskAlignedAssigner select_topk_candidates into a single scatter_add_ by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/25326
• Add Objects365 pretrained YOLO26 models to dataset docs by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/25333
• fix: list TensorRT export arguments by @amanharshx in https://github.com/ultralytics/ultralytics/pull/25338
• Return loss items as name-keyed dicts from criteria by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/25330
• Cache semantic mask bit-depth to avoid reopening every epoch by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/24985
• update ultralytics-inference version to 0.0.30 in documentation by @onuralpszr in https://github.com/ultralytics/ultralytics/pull/25339
• Fix safe_download tar extraction crash with delete=True by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25340
• Make threaded plot threads non-daemon so interpreter shutdown joins them by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25341
• New YOLO26-Depth monocular depth estimation task by @Bovey0809 in https://github.com/ultralytics/ultralytics/pull/25065

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.103...v8.4.104

🌟 Summary

v8.4.103 improves training reliability, result handling, deployment workflows, and documentation—most importantly ensuring warmup finishes on schedule so short training runs reach their intended learning-rate behavior. 🚀

📊 Key Changes

• Training warmup no longer consumes entire short runs 🎯

  • warmup_epochs is now treated as a true epoch count rather than being forced to at least 100 iterations.
  • Warmup is capped so at least the final planned epoch uses the regular learning-rate schedule.
  • Warmup calculations are centralized and remain consistent after batch-size changes, out-of-memory recovery, and timed-training adjustments.
  • Example: a 10-epoch run with 3 warmup epochs now warms up for 3 epochs instead of potentially staying in warmup for the entire run.

• More accurate validation and experiment reporting 📈

  • Confusion matrices now honor the confidence threshold supplied by the validator instead of silently overriding it to 0.25.
  • Validation samples sent to the Platform now include evenly ranked F1 samples across the full image distribution, plus separate best and worst cohorts.
  • Segmentation, pose, and raw confusion-matrix plots are no longer dropped from experiment trackers such as W&B, ClearML, Neptune, and DVC.

• Improved model construction and performance ⚡

  • Fixed backbone width scaling when a layer’s channel count happens to equal the number of classes, preventing unexpectedly oversized models.
  • Accelerated ModelEMA updates with batched PyTorch operations where supported.
  • Made model profiling safer by profiling copied layers rather than modifying shared model modules.

• More robust results and mask processing 🖼️

  • Results methods such as plot(), save_txt(), save_crop(), summary(), and verbose() now work correctly with NumPy-backed results from Results.numpy().
  • Mask processing now binarizes masks before cropping while preserving output behavior.
  • Image-list prediction sources now follow HTTP 308 redirects on supported Python versions.

• Safer and clearer exports 🔧

  • CoreML compatibility checks now run before conversion and reject unsupported dynamic classification and RT-DETR exports earlier.
  • Removed the unsupported Hailo opset argument.
  • RKNN exports now warn users when a requested opset above 19 is automatically reduced.

• Expanded Platform and API capabilities ☁️

  • Platform REST API documentation now covers supported API-key operations across datasets, models, training, exports, deployments, storage, activity, and workspaces.
  • Added documentation for Slack alerts for training, export, and deployment results.
  • Updated Platform On Premise and Docker image references to use rolling latest tags.
  • Linux GPU Docker guidance now recommends CDI --device reservations for more reliable access across host daemon reloads.

• Documentation and workflow improvements 📚

  • YOLO26 documentation now presents inference quickstarts above the fold and points open-vocabulary YOLOE-26 content to its dedicated page.
  • Guides are organized by project stage, from data preparation through deployment and edge hardware.
  • Added contextual “What’s Next” links across prediction, dataset, training, validation, and export pages.
  • Updated broken Tiger-Pose video links and moved YOLOv5 documentation into a clearly labeled legacy section.
  • Updated GitHub Actions workflows to use actions/setup-python@v7.

🎯 Purpose & Impact

• Short training jobs now train as intended ✅ Users no longer risk spending an entire run in warmup with very low learning rates, which should improve optimization and make results more representative of the configured schedule.
• Metrics and diagnostics are more trustworthy 📊 Confidence thresholds, confusion matrices, F1 samples, and tracker plots now better reflect actual validation settings and model behavior.
• Model sizes are more predictable 🧠 Models created from YAML configurations will no longer become unexpectedly large in certain class-count configurations.
• NumPy interoperability is substantially better 🔄 Users can convert results to NumPy and continue using common output, plotting, cropping, and export methods.
• Deployment is safer and easier to troubleshoot 🛠️ Unsupported export combinations fail earlier, while RKNN overrides and Docker GPU requirements are now communicated clearly.
• Platform users get better automation and integration options 🔔 Slack alerts and broader API-key support make it easier to monitor and manage training and deployment programmatically.
• Recommended model: For new projects, use YOLO26, or streamline annotation, training, and deployment with Ultralytics Platform.

What's Changed

• Fix fuzz subprocess source and signatures by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25294
• Modernize Platform links and Docker workflows by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25295
• Align Platform API docs with API-key support by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25296
• Update broken YouTube video links in docs by @RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/25314
• Bump actions/setup-python from v6 to v7 in /.github/workflows by @UltralyticsAssistant in https://github.com/ultralytics/ultralytics/pull/25311
• Honor validation confidence in confusion matrices by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25320
• Document Platform Slack alerts by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25301
• Remove redundant opset argument from Hailo export by @lakshanthad in https://github.com/ultralytics/ultralytics/pull/25305
• Sample validation results across F1 ranks by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25300
• Repair broken forward links in the newcomer training path by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/25253
• Group guides index into categorized sections and split YOLOv5 legacy from the nav by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/25307
• Validate CoreML combinations before export by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25323
• Follow HTTP 308 redirects in autocast_list URL sources by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/25313
• Validate effective final training batch by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25322
• Warn when RKNN export lowers a user-specified opset by @synml in https://github.com/ultralytics/ultralytics/pull/25303
• Faster ModelEMA update with batched _foreach_lerp_ by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/25315
• Move quick start and predict example above the fold on the YOLO26 page by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/25312
• Fix Mask/Pose curves and raw confusion matrix dropped from experiment trackers by @JESUSROYETH in https://github.com/ultralytics/ultralytics/pull/25319
• Binarize masks before cropping in process_mask by @JESUSROYETH in https://github.com/ultralytics/ultralytics/pull/25298
• Fix backbone layers skipping width scaling when their channel count equals nc by @JESUSROYETH in https://github.com/ultralytics/ultralytics/pull/25297
• Fix Results methods crashing on numpy-backed results by @JESUSROYETH in https://github.com/ultralytics/ultralytics/pull/25318
• Prevent training from ending during warmup by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25321

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.102...v8.4.103

🌟 Summary

🚀 Ultralytics 8.4.102 makes MuSGD training up to 8× faster while expanding YOLO26 edge deployment, improving export reliability, and strengthening dataset and GPU error handling.

📊 Key Changes

• ⚡ Faster MuSGD optimizer

  • Batches Newton–Schulz orthogonalization across compatible parameter matrices instead of processing each parameter separately.
  • Uses fused PyTorch foreach operations for momentum and SGD updates.
  • Reduces thousands of small CUDA kernel launches, making MuSGD much less launch-bound.
  • Preserves the existing single-tensor muon_update API while adding an optimized internal batched path.
  • Reported impact: up to 8× faster optimizer steps, particularly for models such as YOLO26n.

• 📱 YOLO26 semantic segmentation on Hailo

  • Adds direct Hailo HEF export and inference for YOLO26 semantic segmentation.
  • Hailo-8/8L devices return logits for host-side upsampling and class reduction.
  • Hailo-10/15 devices can compile multi-class upsampling and ArgMax directly into the accelerator graph, returning compact class maps.
  • Uses task-specific lightweight calibration datasets when no dataset is provided.

• 🛡️ More reliable GPU training recovery

  • Treats CUBLAS_STATUS_ALLOC_FAILED as a recoverable first-epoch memory error, allowing automatic batch-size reduction and retry behavior similar to out-of-memory errors.

• 📦 Improved classification dataset handling

  • Preserves global class IDs for sparse NDJSON train and validation splits.
  • Correctly maps all samples to class 0 when single_cls=True.
  • Uses the actual labels exposed by the dataset for class-count validation.

• 🔧 Safer model export and inference

  • RT-DETR ONNX export now validates that the resolved opset is compatible before export.
  • INT8 classification calibration respects the requested dataset split and falls back cleanly when a test split is unavailable.
  • Augmented inference now checks the model’s final detection head directly, improving compatibility with legacy RT-DETR checkpoints.

• ✅ Clearer validation errors

  • Segmentation now raises an actionable ValueError when mask_ratio is too large for the selected image size, instead of failing later inside OpenCV.

• 📚 Platform and deployment documentation updates

  • Documents optional immutable dataset versions for cloud training, linking trained models to the exact dataset snapshot used.
  • Updates Hailo accuracy-retention guidance for segmentation, pose, OBB, and classification.
  • Reflects the current Ultralytics Platform deployment coverage of 42 regions.

🎯 Purpose & Impact

• Faster training: MuSGD users should see substantially shorter optimizer steps and improved GPU utilization, especially on CUDA workloads with many parameters.
• Broader edge deployment: YOLO26 semantic segmentation can now be deployed directly to supported Hailo accelerators, including Raspberry Pi-oriented hardware.
• More dependable production workflows: Automatic recovery from additional CUDA allocation failures reduces avoidable training interruptions.
• Fewer dataset-related crashes: Correct handling of sparse class metadata and single-class training prevents invalid labels and CUDA assertion failures.
• More predictable exports: Earlier RT-DETR opset validation and improved calibration split handling provide clearer errors and more reliable INT8 export behavior.
• Better experiment reproducibility: Immutable dataset versions in Ultralytics Platform help ensure that models can be traced back to the precise data used for training.

What's Changed

• Document Save Dataset Version for cloud training by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25281
• Fix GPU training failures from production logs by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25284
• Document Hailo per-task INT8 accuracy retention by @JESUSROYETH in https://github.com/ultralytics/ultralytics/pull/25283
• Raise clean error when mask_ratio exceeds imgsz by @darshil929 in https://github.com/ultralytics/ultralytics/pull/25282
• Add Hailo semantic segmentation export and inference for YOLO26 by @JESUSROYETH in https://github.com/ultralytics/ultralytics/pull/25280
• Fix RT-DETR augmented prediction fallback by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25289
• Validate classification calibration split by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25292
• Relabel classification samples for single-class training by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25290
• Validate RT-DETR ONNX opset before export by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25291
• 8x faster MuSGD optimizer step with batched Newton-Schulz by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/25288

New Contributors

• @darshil929 made their first contribution in https://github.com/ultralytics/ultralytics/pull/25282

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.101...v8.4.102

🌟 Summary

🚀 Ultralytics v8.4.101 expands Hailo deployment support across nearly the full YOLO task range and strengthens NDJSON dataset reliability, Platform validation insights, and documentation.

📊 Key Changes

• 🧠 Expanded Hailo export and inference

  • format="hailo" now supports:
    • Object detection
    • Instance segmentation
    • Pose estimation
    • Oriented bounding boxes (OBB)
    • Image classification
  • Added support for YOLOv8 and YOLO11 pose, OBB, and classification models.
  • Added YOLO26 classification export support.
  • Hailo classification models now return probabilities directly using on-chip softmax.
  • Pose and OBB outputs are decoded through the standard Ultralytics inference pipeline.
  • YOLO26 pose, OBB, and segmentation remain unsupported and are rejected before compilation.
  • Pose, OBB, and classification paths were validated on Hailo-8L with HailoRT 4.23 and DFC 3.33. Other targets are accepted but require separate validation.

• 🛡️ More reliable NDJSON dataset conversion

  • Signed NDJSON URLs with authentication query parameters are now recognized correctly.
  • Added file locking to prevent multiple training jobs from converting the same dataset simultaneously.
  • Interrupted downloads can resume without unnecessarily redownloading completed images.
  • Dataset completion markers prevent partially converted datasets from being reused.
  • Preserves safe content hashes in generated filenames while maintaining collision protection.
  • Added filelock as a required dependency.

• 📈 Improved Platform validation reporting

  • Training now records both the worst-performing and best-performing validation images.
  • Reports up to 25,000 images from each cohort, while preserving the existing 50,000-row transport limit.
  • Includes true positives, false positives, and false negatives for each selected image.

• 🏢 Clearer On Premise data boundaries

  • Documentation now explains which data stays local and which information is uploaded.
  • Source dataset pixels remain on the local machine, while labels, annotations, metrics, and the best model checkpoint can be sent to Platform.
  • Documents the exact Docker images used for ARM64, CPU, and NVIDIA environments.

• 🗂️ Improved dataset ingest documentation

  • Documents the classMapping parameter for mapping incoming archive classes to existing dataset classes.
  • Supports numeric class indexes, class names, and null to skip labels.

• 📚 Documentation and maintenance updates

  • Refreshed the VisDrone tutorial video and MNIST creator link.
  • Removed the obsolete HUB-SDK entry from the CI documentation.
  • Added internal NDJSON conversion reference documentation.

🎯 Purpose & Impact

• 🚀 Broader edge deployment options: Hailo users can deploy more YOLO task types directly to Hailo HEF without building separate conversion pipelines.
• ⚡ Better hardware acceleration: Classification softmax runs on the Hailo device, while pose and OBB decoding integrates with normal Ultralytics inference.
• 🔒 Safer concurrent training: Multiple jobs using the same NDJSON source are less likely to corrupt or duplicate cached data.
• ⏱️ Faster recovery from interruptions: Failed image downloads can resume from the existing cache instead of starting over.
• 🔍 More actionable model evaluation: Platform users can inspect both failure cases and strong-performing validation examples, making dataset and model improvements easier.
• 🏢 More transparent enterprise workflows: On Premise users have clearer expectations about data residency and model-weight uploads.
• ⚠️ Deployment note: Hailo compilation is hardware- and compiler-specific. New pose, OBB, and classification exports should be validated on the target accelerator before production use.

What's Changed

• Add https://youtu.be/rTblZN9IRDo to docs by @RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/25268
• docs: document classMapping parameter for dataset ingest by @amanharshx in https://github.com/ultralytics/ultralytics/pull/25272
• Fix MNIST creator link and CI table by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25273
• Document On Premise training data boundaries by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25274
• Fix training from signed NDJSON URLs by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25275
• Add Platform image validation metrics by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25265
• Capture best and worst validation images by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25279
• Fix concurrent NDJSON cache reuse by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25278
• Add Hailo pose, OBB, and classification export and inference for YOLOv8 and YOLO11 by @JESUSROYETH in https://github.com/ultralytics/ultralytics/pull/25276

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.100...v8.4.101

🌟 Summary

🚀 v8.4.100 expands edge deployment capabilities with Hailo instance segmentation support for YOLOv8 and YOLO11, while improving deployment guidance, training validation, Platform workers, fuzz testing, and export reliability.

📊 Key Changes

• 🟢 Hailo instance segmentation export and inference

  • Added Hailo HEF export support for YOLOv8-seg and YOLO11-seg.
  • Hailo exports raw segmentation tensors, while Ultralytics handles decoding, bounding-box processing, NMS, and mask generation during inference.
  • Segmentation models do not use Hailo’s on-chip NMS configuration.
  • YOLO26 segmentation is explicitly rejected with a clear error because Hailo-8/8L hardware cannot allocate its attention-based architecture.

• 📚 Improved Hailo INT8 deployment documentation

  • Added practical accuracy expectations for YOLOv8, YOLO11, and YOLO26 on Hailo-8L.
  • Documents that calibration data should closely match the production domain.
  • Notes that YOLO26 confidence scores may be approximately 0.05 lower on Hailo devices.
  • Explains why attention layers can be more sensitive to INT8 quantization on Hailo-8/8L.

• ⚠️ Clearer TensorRT compatibility guidance

  • Platform documentation now warns that downloaded TensorRT engines must match the target GPU family, TensorRT version, CUDA runtime, and build environment.
  • Users with mismatched environments are directed to export engines locally on the deployment device.

• 🧪 Broader and more effective fuzz testing

  • Fuzzing now covers tracking, cached video, malformed media, Unicode paths, additional model families, and CoreML on macOS.
  • Duplicate commands are skipped and shell-safe reproducer commands are generated.
  • Fuzzing budgets increase from 285 to 300 minutes per run.

• 🛑 Cleaner training errors for invalid small configurations

  • Training now raises an actionable error when batch=1 is used with an image size too small for BatchNorm.
  • This replaces a confusing low-level PyTorch failure with guidance to increase the batch size or image size.

• 🔧 More reliable Axelera exports

  • Pins omnimalloc==0.5.0 to avoid compatibility failures with Axelera Devkit 1.7.0.
  • Runtime checks now validate the same allocator version before export.

• ☁️ Improved Platform managed-worker integration

  • Reuses existing Platform training callbacks for managed workers.
  • Supports configurable Platform API endpoints through PLATFORM_API_URL.
  • Training output paths now respect the current SETTINGS["runs_dir"] value at runtime.

🎯 Purpose & Impact

• Edge users gain a new deployment option: YOLOv8 and YOLO11 segmentation models can now run on Hailo accelerators, including Hailo-8, Hailo-8L, Hailo-10, and Hailo-15.
• Inference behavior remains consistent: Hailo segmentation outputs are converted into the standard Ultralytics format, allowing existing predictor post-processing and mask workflows to remain unchanged.
• Deployment accuracy expectations are clearer: Users can better estimate INT8 performance and avoid misleading comparisons between PyTorch and Hailo results.
• Fewer deployment surprises: TensorRT and Axelera compatibility warnings help users identify environment mismatches before runtime failures.
• More understandable training failures: Invalid batch/image-size combinations now fail early with a practical fix instead of producing an obscure internal exception.
• Higher software reliability: Expanded fuzz coverage and duplicate elimination should uncover more real edge cases while reducing wasted test time.

📌 For the simplest workflow to annotate datasets, train models, and deploy them, visit the Ultralytics Platform.

What's Changed

• Document Hailo INT8 accuracy expectations and deployment guidance by @JESUSROYETH in https://github.com/ultralytics/ultralytics/pull/25260
• Warn on TensorRT/CUDA version mismatch in Platform Models export docs by @lakshanthad in https://github.com/ultralytics/ultralytics/pull/25263
• Expand fuzz exploration envelope by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25262
• Raise clean error for batch=1 training at minimum imgsz by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25264
• Fix flaky Axelera export environment by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25267
• Reuse Platform callbacks for managed workers by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25266
• Add Hailo segmentation export and inference for YOLOv8 and YOLO11 by @JESUSROYETH in https://github.com/ultralytics/ultralytics/pull/25259

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.99...v8.4.100

🌟 Summary

🚀 v8.4.99 adds native Hailo inference support, allowing Ultralytics-exported Hailo models to run directly with predict and val, while also improving training memory usage, platform checkpoint handling, downloads, and compatibility.

📊 Key Changes

• 🧩 New Hailo inference backend

  • Adds lazy HailoRT support for Ultralytics-exported HEF model directories.
  • Supports YOLOv8, YOLO11, and YOLO26 detection outputs.
  • Automatically converts Hailo NMS results and decodes YOLO26’s NMS-free outputs.
  • Enables commands such as:

    from ultralytics import YOLO
    
    model = YOLO("yolo11n_hailo_model")
    results = model.predict("image.jpg")
    

  • Adds Hailo backend documentation and removes outdated warnings and manual post-processing instructions.
  • Requires HailoRT on the target device and the exported directory’s metadata.yaml file.

• 🧠 Lower training memory usage

  • Reduces semantic segmentation target construction memory substantially, with measured temporary usage dropping from about 525 MB to 136 MB.
  • Optimizes dense-dataset target assignment, reducing peak assigner memory from about 1.52 GB to 0.89 GB in the reported test.
  • These changes can make training large or densely annotated datasets, such as Objects365, more practical on GPUs with limited VRAM.

• ☁️ More reliable Platform checkpoint uploads

  • Successful periodic checkpoint uploads are now recorded by Ultralytics Platform.
  • Uploads remain asynchronous during training, while the final upload continues to wait safely for earlier uploads.
  • Checkpoints are associated with the correct server-issued training run, helping Platform maintain the canonical best.pt.

• 🔗 Improved download and redirect handling

  • Uses the existing requests dependency for URL checks and downloads.
  • HTTP 308 redirects now work reliably on older Python versions, including Python 3.8–3.10, without adding a custom urllib compatibility layer.

• 🐍 Better Python 3.13 checkpoint compatibility

  • Restricted checkpoint loading now recognizes both native and stable public pathlib path names.
  • This prevents failures when checkpoints are created under one Python version or operating system and loaded under another.

• 🛠️ CI and documentation maintenance

  • All workflows now use Ultralytics’ shared retried setup-uv action.
  • Dataset license URLs are deduplicated through shared metadata.
  • Corrects LVIS class 666 from manager/through to the accurate manger/trough.
  • Makes exporter output show prediction and validation commands only for formats with supported inference backends.

🎯 Purpose & Impact

• Hailo users can now use a more familiar Ultralytics workflow for exported edge models instead of manually integrating HailoRT output parsing. This simplifies deployment on Hailo hardware, although HailoRT installation and compatible hardware are still required.
• YOLO26 deployment is more streamlined, as its one-to-one, NMS-free outputs are decoded automatically by the Hailo backend.
• GPU training may require less VRAM, particularly for segmentation and dense detection datasets, enabling larger batch sizes or higher-resolution training in some environments.
• Ultralytics Platform training runs become more dependable, with fewer risks of stale or untracked checkpoint uploads.
• Downloads and model loading are more cross-version compatible, improving reliability for users on older Python releases and Python 3.13.
• Overall, this release strengthens deployment, training efficiency, and infrastructure reliability without changing existing model outputs.

What's Changed

• Fix Hailo post-export guidance by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25245
• Deduplicate canonical dataset license URLs by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/25249
• Fix misspelled LVIS class 666 to manger/trough by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/25233
• Use shared setup-uv action everywhere by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25251
• Reduce semantic segmentation target memory in v8SegmentationLoss by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/25244
• Reduce TaskAlignedAssigner peak memory on dense datasets by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/25240
• Fix HTTP 308 redirect handling in safe_download on Python <=3.10 by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/25252
• Revert HTTP 308 urllib backport by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25256
• Fix stale buffer lists in the profile_ops memory-envelope comment by @JESUSROYETH in https://github.com/ultralytics/ultralytics/pull/25235
• Use requests for redirecting downloads by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25257
• Record successful Platform checkpoint uploads by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25255
• Fix Python 3.13 pathlib checkpoint loading by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25258
• Add Hailo inference backend by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25247

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.98...v8.4.99

🌟 Summary

v8.4.98 improves Hailo export compatibility for legacy YOLO checkpoints while preserving strict model-family validation. 🚀

📊 Key Changes

• Legacy checkpoint support for Hailo export 🛠️
  Hailo exports can now identify the model family from its architecture when older checkpoints do not include the optional yaml_file metadata.

• Topology-based model detection 🔍
  The exporter examines backbone and head components to recognize supported YOLOv8, YOLO11, and YOLO26 checkpoints when metadata is missing.

• Existing validation remains intact ✅
  The detection-head checks and supported-family restrictions were preserved, preventing unsupported architectures such as YOLOv5, YOLOv10, and YOLO12 from being sent to Hailo’s Dataflow Compiler.

• Patch version updated 📦
  The package version was incremented from 8.4.97 to 8.4.98.

🎯 Purpose & Impact

• More reliable deployment of older models: Legacy trained checkpoints can now be exported to Hailo without requiring metadata that was not consistently stored in older versions.
• No change to supported model scope: The update improves checkpoint recognition rather than adding unsupported model families.
• Safer exports: Existing architecture validation continues to catch incompatible models before Hailo compilation, reducing confusing downstream failures.
• Broad compatibility coverage: Testing confirmed that metadata-stripped YOLOv8, YOLO11, and YOLO26 checkpoints are accepted, while unsupported families are correctly rejected. 🧪

What's Changed

• Support legacy checkpoints in Hailo export by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25237

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.97...v8.4.98

🌟 Summary

🚀 Ultralytics 8.4.97 adds direct Hailo HEF export, making it much easier to compile YOLOv8, YOLO11, and YOLO26 detection models for Hailo edge accelerators while also delivering several export fixes, inference improvements, and documentation updates.

📊 Key Changes

• Direct Hailo HEF export

  • Added model.export(format="hailo") and the equivalent CLI command.
  • Supports Hailo-8, Hailo-8L, Hailo-10H, Hailo-15H, and Hailo-15L.
  • Supports standard YOLOv8, YOLO11, and YOLO26 detection models, including custom detection models.
  • Automatically generates parser settings, calibration data, HailoRT post-processing, metadata, and the final .hef file.
  • YOLOv8 and YOLO11 exports use HailoRT NMS, while YOLO26 retains its NMS-free output path.
  • Replaces the previous lengthy manual ONNX-to-DFC workflow with Python, CLI, and Ultralytics Platform workflows.
  • Hailo exports are INT8-only and recommend at least 1,024 representative calibration images for production accuracy.

  from ultralytics import YOLO
  
  model = YOLO("yolo11n.pt")
  model.export(format="hailo", name="hailo8l", data="my_dataset.yaml")
  

• CoreML export reliability improvements

  • Embedded CoreML NMS is now restricted to object detection models.
  • Segmentation and pose exports automatically disable incompatible NMS settings with a warning.
  • CoreML NMS exports now preserve class names and other model metadata.
  • NMS-based ML Program exports default to FP16 when no precision is specified, helping maintain Xcode Preview and Neural Engine compatibility.

• Inference and export bug fixes

  • Fixed TorchScript segmentation inference crashes caused by tensors being split between CPU and GPU.
  • save_crop=True now reliably writes crops even when save, save_txt, and show are disabled.
  • Reduced GPU-to-CPU synchronization overhead in Results.plot(), improving visualization performance for detection, segmentation, and pose results.
  • Restored the user-selected torch.compile() mode, including max-autotune.
  • Fixed custom PyTorch modules whose fuse() method does not accept Ultralytics-specific arguments.
  • INT8 calibration now correctly validates the requested dataset split.
  • Albumentations spatial transforms now correctly transform semantic masks.
  • YOLOE-Seg visual-prompt training no longer crashes when using end-to-end loss settings.
  • Dataset downloads now respect HTTP_PROXY and HTTPS_PROXY environment variables.

• Deployment and ecosystem updates

  • LiteRT is now used in new export examples, while the historical TFLite guide is clearly marked deprecated.
  • Added ROS2 rclpy examples alongside the existing ROS1 workflow.
  • Updated Jetson TensorRT documentation with device-specific export targets and clearer engine portability warnings.
  • Added verified creator and license metadata across dataset documentation.
  • Refreshed the documentation homepage around common user goals, including training, prediction, deployment, tracking, and the Ultralytics Platform.
  • Updated OpenVINO benchmark references to version 2026.2.1.
  • Updated Slack GitHub Actions used by release and CI workflows.

🎯 Purpose & Impact

• ⚡ Simpler Hailo deployment: Users can now move from a trained YOLO model to a deployable Hailo HEF with one supported export command instead of manually managing ONNX graphs, end nodes, model scripts, and NMS configuration.
• 🎯 Better edge accuracy: Automatic calibration handling and the recommendation to use representative datasets help reduce accuracy loss from INT8 quantization.
• 🧩 Broader hardware coverage: Direct export supports multiple Hailo accelerator generations, while Ultralytics Platform provides a convenient alternative when local Hailo compiler dependencies are unavailable.
• 🍎 More dependable CoreML models: Detection exports retain NMS behavior and metadata, while non-detection models avoid unsupported pipelines.
• 🚄 Faster visualization: Results.plot() now performs fewer per-box and per-keypoint device synchronizations, which can improve rendering performance on GPU workloads.
• 🛠️ More predictable workflows: Fixes to calibration splits, proxy-aware downloads, TorchScript devices, crop saving, and custom module fusion address failures that could otherwise be confusing or silent.
• 📚 Improved accessibility: Updated documentation makes it easier for both new and experienced users to train, export, and deploy Ultralytics models across Hailo, Jetson, ROS2, LiteRT, CoreML, and other platforms.

What's Changed

• Fix CoreML NMS for non-detect models by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25198
• Migrate remaining TFLite examples to LiteRT by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25208
• Batch PR review findings before final gate by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25207
• Fix #25205: trust environment in aiohttp.ClientSession and allow it … by @muzhikov in https://github.com/ultralytics/ultralytics/pull/25206
• Fix save_crop=True silently writing no crops without save/save_txt/show by @JESUSROYETH in https://github.com/ultralytics/ultralytics/pull/25204
• Fix device mismatch for TorchScript segment inference on GPU by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/25201
• Fix TVPDetectLoss misnamed box_loss — v8DetectionLoss index 1 is cls, not box by @Zenka737 in https://github.com/ultralytics/ultralytics/pull/25197
• Avoid setting compute_precision for CoreML NMS export by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/25200
• Fix spatial Albumentations transforms ignoring semantic_mask by @ahmet-f-gumustas in https://github.com/ultralytics/ultralytics/pull/25191
• Fix metadata loss on CoreML export with NMS by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/25199
• Bump slackapi/slack-github-action from 3.0.5 to 4.0.0 in /.github/workflows by @dependabot[bot] in https://github.com/ultralytics/ultralytics/pull/25210
• Update OpenVINO benchmarks version to 2026.2.1 by @lakshanthad in https://github.com/ultralytics/ultralytics/pull/25211
• Fix custom PyTorch module fusion by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25215
• Add verified dataset license metadata by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25214
• fix formatting in C++ README for YOLO model references by @onuralpszr in https://github.com/ultralytics/ultralytics/pull/25221
• Rebuild docs landing page as a goal-based router by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/25217
• Fix TVPSegmentLoss crashing end-to-end YOLOE-Seg visual-prompt training on tal_topk2 by @JESUSROYETH in https://github.com/ultralytics/ultralytics/pull/25209
• Update C++ README files to specify Ultralytics in model references by @onuralpszr in https://github.com/ultralytics/ultralytics/pull/25222
• Validate INT8 calibration dataset split by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25223
• Stop reporting intentional fuzz rejections by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25224
• Bold what/how emphasis in the docs organization bullets by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/25226
• Document Platform Jetson export targets by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25232
• Avoid per-box and per-keypoint GPU syncs in Results.plot() by @JESUSROYETH in https://github.com/ultralytics/ultralytics/pull/25230
• Restore max-autotune option for torch.compile() by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/25229
• Extend ros-quickstart guide with ROS2 (rclpy) examples by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/25169
• Add direct Hailo HEF export by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25231

New Contributors

• @muzhikov made their first contribution in https://github.com/ultralytics/ultralytics/pull/25206

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.96...v8.4.97

🌟 Summary

Ultralytics v8.4.96 strengthens NDJSON dataset security and dramatically improves conversion speed, while adding YOLO26 depth deployment guidance, On Premise Platform documentation, and numerous reliability fixes. 🚀

📊 Key Changes

• 🔐 Faster and safer NDJSON conversion

  • Reverts the previous overbuilt validation approach and validates records during the existing hash scan.
  • Generates collision-free filenames from stable record indexes instead of trusting potentially unsafe source names.
  • Uses dense numeric class directories for classification datasets, while preserving original class names through a hidden metadata map.
  • Removes the extra full-dataset validation pass and avoids unnecessary path-object allocations.
  • Reported performance improves by approximately 7.8× for detection and 6.8× for classification compared with the reverted implementation.

• 📏 YOLO26 depth support in Rust inference

  • Adds depth estimation as a supported task with yolo26n-depth.onnx.
  • Supports depth visualization options such as --colormap and --depth-viz.
  • Adds Rust API examples for reading per-pixel depth maps, valid ranges, and center-pixel depth values.
  • Updates ultralytics-inference documentation to version 0.0.29.

• 🚀 New Intel deployment guide

  • Documents YOLO26 deployment with Intel DL Streamer, OpenVINO, GStreamer, and Docker.
  • Includes FP32, FP16, and INT8 export examples for Intel CPUs, GPUs, and NPUs.
  • Covers GPU/NPU inference, video output, and multi-stream processing.

• 🏢 Expanded Ultralytics Platform On Premise documentation

  • Explains how Enterprise users can connect Linux, macOS, or Windows workers.
  • Documents local dataset browsing, annotation, training, and model storage while keeping image and video pixels on the customer’s computer.
  • Adds plan metadata across Platform documentation and clarifies Enterprise capabilities.

• 🧩 Reliability and correctness fixes

  • Results.update(probs=...) now correctly wraps probabilities in Probs, restoring methods such as summary(), plot(), and exports.
  • RT-DETR standalone validation now honors single_cls and classes.
  • Keypoints and segmentation polygons remain aligned with boxes after letterbox resizing.
  • Classification confusion matrices no longer include an irrelevant background class.
  • FastSAM no longer overwrites the first detection when handling full-image boxes.
  • Segmentation loss no longer silently skips images when instances are fewer than the batch size.
  • save_one_box() now returns crops in the requested format regardless of whether the crop is saved.
  • End-to-end NMS now correctly returns original detection indexes when return_idxs=True.
  • Tracker motion compensation, trainer recovery logic, TensorBoard channel handling, YOLOE visual-prompt training, and configuration validation receive targeted fixes.

• 📚 Documentation improvements

  • Adds clearer semantic segmentation dataset guidance, including the masks/ folder fallback behavior.
  • Corrects augmentation and export argument tables and updates the tracker default to tracktrack.yaml.
  • Labels many documentation examples as text for improved rendering and readability.
  • Updates CoreML and LiteRT mobile benchmarks to include YOLO26 Depth results.

🎯 Purpose & Impact

• Safer dataset processing: Untrusted NDJSON filenames and class names can no longer control generated filesystem paths, reducing traversal and collision risks. 🛡️
• Faster large-dataset workflows: NDJSON conversion avoids redundant scans and scales much better for millions of records.
• Broader YOLO26 deployment: Users can now more easily run depth estimation on Rust, Apple, Android, and Intel hardware.
• Better production flexibility: Enterprise customers can train through the Ultralytics Platform while keeping sensitive data and compute on premise.
• More dependable APIs and results: Several fixes prevent confusing runtime errors, incorrect metrics, misaligned outputs, and silently dropped training data.
• Clearer documentation: Updated examples, defaults, benchmarks, and platform plan labels make setup and deployment easier for both new and experienced users. ✅

What's Changed

• Fix augmentation-args and export-args doc macro cells by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/25163
• Document masks/-folder fallback in semantic task faq by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/25161
• Fix Results.update(probs=...) to wrap probs in Probs like the other attributes by @JESUSROYETH in https://github.com/ultralytics/ultralytics/pull/25157
• Fix RT-DETR standalone val ignoring single_cls and classes by @JESUSROYETH in https://github.com/ultralytics/ultralytics/pull/25158
• Fix scale_coords letterbox pad so keypoints and mask polygons stay aligned with boxes by @JESUSROYETH in https://github.com/ultralytics/ultralytics/pull/25156
• Fix classify ConfusionMatrix gaining a phantom background class by @JESUSROYETH in https://github.com/ultralytics/ultralytics/pull/25155
• Fix stale tracker default and dropped row in object-blurring solutions table by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/25170
• Document Enterprise On Premise workers by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25167
• Use dataset channels for TensorBoard graph tracing by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/25160
• Forward validation compile mode to attempt_compile by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/25162
• Validate CopyPaste mode explicitly by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/25168
• Add guide for deploying Ultralytics YOLO26 on Intel Core Ultra Series3 using DL Streamer and OpenVINO. by @tbujewsk in https://github.com/ultralytics/ultralytics/pull/25078
• Rewrite On Premise docs for customers by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25174
• Fix v8SegmentationLoss silently dropping images when overlap_mask=False and instances < batch size by @JESUSROYETH in https://github.com/ultralytics/ultralytics/pull/25173
• Fix save_one_box to return the crop in the requested BGR format regardless of save by @JESUSROYETH in https://github.com/ultralytics/ultralytics/pull/25172
• Update CoreML and LiteRT mobile depth benchmarks by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25177
• Improve Platform code example readability by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25179
• Remove obsolete docs workflow badge by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25180
• Fix FastSAM postprocess overwriting the first box by @JESUSROYETH in https://github.com/ultralytics/ultralytics/pull/25175
• Fix ECC camera-motion translation not rescaled after downscale in GMC by @JESUSROYETH in https://github.com/ultralytics/ultralytics/pull/25181
• Fix best_fitness falsy check and NaN recovery dead-code branches in trainer by @Zenka737 in https://github.com/ultralytics/ultralytics/pull/25176
• Fix YOLOESegModel.loss mixing text prompts into visual-prompt training by @JESUSROYETH in https://github.com/ultralytics/ultralytics/pull/25178
• Validate CLI choice arguments in config by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25182
• Fix stale package imports in Python subprocesses by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/24447
• Revert overbuilt subprocess import fix by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25190
• Update ultralytics-inference version to 0.0.28 in rust docs by @onuralpszr in https://github.com/ultralytics/ultralytics/pull/25192
• Document the depth task in the Rust inference docs by @onuralpszr in https://github.com/ultralytics/ultralytics/pull/25193
• fix: validate NDJSON image and split output paths by @amanharshx in https://github.com/ultralytics/ultralytics/pull/25188
• Fix non_max_suppression dropping return_idxs for end-to-end models and remove dead ConfusionMatrix.matrix() method by @Zenka737 in https://github.com/ultralytics/ultralytics/pull/25187
• Fix uint8 overflow in polygons2masks_overlap dropping instances past 128 overlaps by @JESUSROYETH in https://github.com/ultralytics/ultralytics/pull/25185
• Update rust docs for the jet and disparity depth defaults and version 0.0.29 by @onuralpszr in https://github.com/ultralytics/ultralytics/pull/25195
• Revert #25188 and secure NDJSON paths without a prepass by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25196

New Contributors

• @tbujewsk made their first contribution in https://github.com/ultralytics/ultralytics/pull/25078
• @Zenka737 made their first contribution in https://github.com/ultralytics/ultralytics/pull/25176

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.95...v8.4.96

🌟 Summary

v8.4.95 improves training reliability, RT-DETR detection limits, input compatibility, and dataset usability—headlined by safer checkpoint loading for Platform GPU jobs. 🚀

📊 Key Changes

• 🔒 Fixed restricted training checkpoint loading — Raw checkpoints produced by Platform GPU workers no longer include training-only loss objects or NumPy scalar metadata that can be rejected by secure weights_only=True loading.
• 🎯 Improved RT-DETR max_det support — The detection limit is now correctly applied during native prediction, validation, and model exports, including CoreML and other deployment formats.
• 🖼️ Expanded NumPy input compatibility — NumPy images with grayscale, gray-plus-alpha, BGR, or BGRA channels are now normalized to match the model’s expected channel count.
• 🔢 Fixed uint8 tensor preprocessing — Integer image tensors are handled without invalid floating-point dtype operations, while existing floating-point normalization behavior is preserved.
• 📦 Reduced automatic COCO downloads — COCO detection, segmentation, and pose datasets no longer download the optional 7 GB test2017.zip by default. Standard downloads are now approximately 20.2–20.3 GB instead of about 27 GB.
• 🧪 Made segmentation conversion deterministic — Mask files are processed in sorted filename order, improving repeatability across runs.
• 📈 Corrected pose metric curve labels — Removed duplicate box labels so pose metrics align correctly with dashboards such as Weights & Biases.
• 📝 Clarified prediction and embedding return types — Documentation and type hints now accurately indicate that streaming returns an iterator, while non-streaming calls return lists.
• 🛠️ Simplified edge-case handling — Empty segmentation inputs now return a valid zero box immediately.
• 📚 Updated examples and training documentation — The ObjectCounter example now uses its returned results object, and automatic dataset download behavior is described more precisely.

🎯 Purpose & Impact

• More dependable cloud training: Platform GPU workers can save, resume, and load raw training checkpoints more reliably, reducing avoidable job failures. ✅
• More predictable deployment: RT-DETR exports and validation now respect max_det, helping control output size, latency, and downstream processing requirements.
• Fewer input-related errors: Applications using NumPy arrays or uint8 tensors can handle a wider range of camera and image formats without manual channel conversion.
• Faster, lighter dataset setup: Most COCO users save roughly 7 GB of storage and download time because test images without public ground truth are excluded.
• More reproducible data preparation: Deterministic mask ordering makes conversion results and debugging more consistent.
• Clearer analytics: Correct pose labels prevent mislabeled or missing metric panels in experiment tracking tools.
• No major model architecture change: This release focuses on reliability, compatibility, export behavior, and documentation rather than introducing a new model family. YOLO26 remains the latest recommended Ultralytics model for general use. 🌟

What's Changed

• Stop coco.yaml and coco-pose.yaml from downloading test2017.zip by default by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/25136
• Use ObjectCounter results in the usage example by @parthivdholaria in https://github.com/ultralytics/ultralytics/pull/23417
• Process segmentation masks in deterministic order by @AlexRTer in https://github.com/ultralytics/ultralytics/pull/24390
• Simplify empty segment handling in segment2box by @1548093028 in https://github.com/ultralytics/ultralytics/pull/23770
• Normalize uint8 tensor sources without dtype errors by @jahsef in https://github.com/ultralytics/ultralytics/pull/23743
• Clarify automatic dataset downloads in training docs by @vedantparnaik in https://github.com/ultralytics/ultralytics/pull/24155
• Clarify streamed prediction return types by @thiliapr in https://github.com/ultralytics/ultralytics/pull/24391
• Clarify streamed embedding return types by @waketzheng in https://github.com/ultralytics/ultralytics/pull/24134
• Fix PoseMetrics.curves returning duplicated box labels by @JESUSROYETH in https://github.com/ultralytics/ultralytics/pull/25148
• Normalize NumPy channel count to the model in LoadPilAndNumpy by @JESUSROYETH in https://github.com/ultralytics/ultralytics/pull/25151
• fix(rtdetr): honor max_det in decoder exports and validation by @Hasnaathussain in https://github.com/ultralytics/ultralytics/pull/25153
• Fix restricted loading of training checkpoints by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25154

New Contributors

• @thiliapr made their first contribution in https://github.com/ultralytics/ultralytics/pull/24391
• @parthivdholaria made their first contribution in https://github.com/ultralytics/ultralytics/pull/23417
• @jahsef made their first contribution in https://github.com/ultralytics/ultralytics/pull/23743
• @vedantparnaik made their first contribution in https://github.com/ultralytics/ultralytics/pull/24155
• @1548093028 made their first contribution in https://github.com/ultralytics/ultralytics/pull/23770
• @AlexRTer made their first contribution in https://github.com/ultralytics/ultralytics/pull/24390
• @waketzheng made their first contribution in https://github.com/ultralytics/ultralytics/pull/24134
• @Hasnaathussain made their first contribution in https://github.com/ultralytics/ultralytics/pull/25153

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.94...v8.4.95

🌟 Summary

🚀 Ultralytics 8.4.94 modernizes deployment, improves validation and tracking reliability, and expands edge-device and dataset documentation.

📊 Key Changes

• TorchScript mobile optimization removed 🛠️

  • Removed the deprecated PyTorch Mobile/XNNPACK optimization path and the TorchScript optimize argument.
  • Regular TorchScript export remains available for legacy C++ and LibTorch applications.
  • Mobile deployments are now directed to ExecuTorch, PyTorch’s supported successor for mobile and edge inference.
  • The optimize argument is retained for DEEPX exports only.

• New Ambarella CVflow deployment guide 📷

  • Added a preview workflow for deploying YOLO models on Ambarella SoCs such as CV72 and CV75.
  • Documents SpongeTorch compression-aware training, ONNX export, offline CVflow compilation, AmbaPB inference, and host-side validation.
  • The guide is clearly marked as an early preview and requires Ambarella’s proprietary SDK and toolchain.

• Improved semantic segmentation dataset support 🧩

  • Expanded and corrected documentation for ADE20K, Cityscapes, and Cityscapes8.
  • Clarified manual download requirements, dataset sizes, test-set limitations, licensing, class mappings, and benchmark usage.
  • Semantic segmentation can now automatically detect a root-level masks/ directory when selecting PNG-mask data.
  • Added guidance for using polygon labels and Smart annotation through Ultralytics Platform.

• Dataset configuration validation strengthened ✅

  • Dataset fraction must now be greater than zero, preventing empty training datasets caused by fraction=0.
  • Other fractional settings, such as dropout and scale, continue to allow zero.

• Pose26 training loss corrected 🎯

  • Fixed loss selection when Pose26 models run with end-to-end mode disabled.
  • This ensures keypoint decoding and training loss remain aligned across Pose and Pose26 model heads.

• ReID tracking made more robust 🔍

  • Invalid or empty detection crops are now skipped instead of causing crashes.
  • Original detection ordering is preserved, with missing appearance features represented as None so trackers can fall back to motion or IoU association.

• LVIS downloads reduced 💾

  • Removed the unused 7 GB test2017.zip download because LVIS does not define a test split.
  • Updated the documented total download size to 20.7 GB.

• Documentation and CI cleanup 📚

  • Added an Ambarella integration entry and refreshed deployment guidance.
  • Removed outdated CI table entries and redundant YAML terminology.
  • Added more dataset provenance, annotation, licensing, and usage notes.

🎯 Purpose & Impact

• 📱 More future-proof mobile deployment: Users targeting new mobile or edge applications should migrate from TorchScript Mobile to ExecuTorch. Existing C++ TorchScript workflows continue to work, but mobile optimization through optimize=True is no longer supported.
• ⚙️ Cleaner export behavior: Removing obsolete mobile branches and guards reduces maintenance complexity and avoids relying on deprecated PyTorch APIs.
• 🚀 Broader edge hardware coverage: Ambarella users now have a documented path from YOLO training to compressed, compiled CVflow models, although the integration remains experimental and vendor verification is pending.
• 🛡️ Fewer runtime failures: Invalid dataset fractions, incorrect Pose26 loss selection, and empty ReID crops are detected or handled more safely.
• 📉 Lower setup overhead for LVIS: Fresh LVIS users avoid downloading an unnecessary 7 GB archive.
• 📖 Clearer, more accurate documentation: Dataset limitations, licensing requirements, manual downloads, and semantic segmentation behavior are now easier to understand and less likely to cause setup or evaluation mistakes.

What's Changed

• Add Ambarella export guide by @lakshanthad in https://github.com/ultralytics/ultralytics/pull/25077
• Remove stale CI table entries by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25137
• Reject zero dataset fractions by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25141
• Refactor semantic dataset docs by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/25134
• Fix Pose26 loss selection when end2end=False by @synml in https://github.com/ultralytics/ultralytics/pull/25132
• Fix ReID crash for invalid custom detection crops by @anshulll19 in https://github.com/ultralytics/ultralytics/pull/25135
• Add additional notes to dog-pose, hand keypoints, and tiger-pose entries in pose dataset index by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/25144
• Stop lvis.yaml from downloading the unused test2017.zip archive by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/25143
• Drop incorrect YAML expansion from dataset docs by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/25142
• Remove TorchScript mobile optimization by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25140

New Contributors

• @anshulll19 made their first contribution in https://github.com/ultralytics/ultralytics/pull/25135
• @synml made their first contribution in https://github.com/ultralytics/ultralytics/pull/25132

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.93...v8.4.94

🌟 Summary

Ultralytics v8.4.93 expands Alibaba MNN export capabilities, improves training and inference reliability, and adds major dataset, cloud-storage, and integration documentation updates. 🚀

📊 Key Changes

• Alibaba MNN exports now support dynamic shapes and embedded NMS 🎯

  • dynamic=True enables variable image dimensions for MNN models.
  • nms=True is supported for detect and pose models with static shapes.
  • Unsupported combinations now fail clearly:
    • Dynamic shapes with NMS
    • NMS for segmentation and OBB models
  • Empty NMS outputs are handled safely only when the exported metadata confirms NMS is enabled.
  • Real MNN 3.5.0 inference was validated with YOLO11 detection, pose, and OBB workflows.

• Improved model tracking and vision workflows 👁️

  • Fixed native ReID crashes in BoT-SORT when users provide detections directly without feature tensors.
  • Polygon-based ObjectCounter now determines IN/OUT direction from object motion rather than region shape, improving counting in square and wide regions.
  • YOLOE visual prompts now work with PyTorch tensor inputs.
  • embed=[] no longer causes prediction to fail.

• More reliable training, benchmarking, and export behavior 🛠️

  • Semantic segmentation loss remains finite when an entire batch contains ignore pixels.
  • Legacy benchmark(half=True) and benchmark(int8=True) flags now correctly map to quantized benchmarking instead of silently using FP32.
  • Invalid configuration values such as nbs=0, max_det=0, mask_ratio=0, or negative seeds are rejected early with clear errors.
  • TorchScript optimization now reports a helpful message when XNNPACK is unavailable.
  • CPU torch.compile gracefully falls back to eager execution when no C++ compiler is available.
  • CoreML exports use a safer NumPy compatibility limit to avoid NumPy 2.4 crashes.
  • Axelera exports reject image sizes below the backend’s 64-pixel minimum.

• Improved segmentation and conversion utilities ⚡

  • segment2box is significantly faster while preserving its output.
  • process_mask(upsample=True) now upsamples before cropping, preventing mask pixels from leaking outside bounding boxes.
  • Segment-mask conversion now creates missing output directories and handles Windows grayscale image shapes correctly.
  • Deprecated crop_fraction usage now emits a warning instead of stopping classification transforms.

• Expanded pose evaluation customization 🧍

  • Pose datasets can define custom kpt_oks_sigmas values in their YAML files.
  • Values are validated against the configured number of keypoints and must be positive.

• Ultralytics Platform cloud-storage integrations documented ☁️

  • Added guides for connecting Google Cloud Storage, Amazon S3, and Azure Blob Storage.
  • Platform can index images and YOLO labels in place without copying the source data.
  • Documentation covers read-only credentials, dataset creation, revision pinning, import limits, retries, and current limitations.
  • These integrations require a Pro or Enterprise plan.

• Documentation and deployment updates 📚

  • Added documentation for Apple Core AI and its future .aimodel workflow. Core AI export is not yet available; Core ML remains the supported Apple deployment path.
  • Dataset guides now document using .txt image lists for train, validation, and test splits.
  • Segmentation dataset documentation was refreshed with corrected split counts, class information, download sizes, and clearer usage guidance.
  • Rust inference examples now use ultralytics-inference 0.0.27.
  • Python Docker images now use Python 3.13 on Debian Trixie.
  • Environment variables accept common truthy values such as 1, yes, on, y, and t.

🎯 Purpose & Impact

• Broader deployment support: MNN users can target applications that need variable input sizes or built-in post-processing, while clear restrictions prevent unsupported exports from failing later.
• Fewer runtime crashes: Tracking, segmentation training, tensor-based YOLOE inference, compilation, and export workflows now handle common edge cases more safely.
• More trustworthy results: Fixes to mask cropping, object-counting direction, semantic loss, and benchmark precision ensure outputs and metrics better reflect user settings.
• Faster development cycles: The optimized segment2box implementation and improved small-dataset documentation make testing and experimentation more efficient.
• Simpler cloud workflows: Platform users can train on existing cloud datasets without manually downloading and re-uploading large collections.
• Clearer configuration errors: Invalid arguments are detected at the configuration layer with actionable messages instead of obscure deep runtime tracebacks.
• Apple deployment guidance is clearer: Users can continue using supported Core ML exports today while understanding that Core AI support is planned for a future release. 🚀

What's Changed

• Bump slackapi/slack-github-action from v3.0.3 to v3.0.5 in /.github/workflows by @UltralyticsAssistant in https://github.com/ultralytics/ultralytics/pull/25110
• Add Platform cloud storage integration docs (GCS, S3, Azure) by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25112
• Split Platform cloud storage docs into per-vendor pages by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25113
• Fix ObjectCounter polygon direction to follow object motion by @JESUSROYETH in https://github.com/ultralytics/ultralytics/pull/25104
• Fix BOTSORT native-ReID crash with user-supplied detections by @JESUSROYETH in https://github.com/ultralytics/ultralytics/pull/25102
• Pin numpy<=2.3.5 for CoreML export to avoid numpy>=2.4 crash by @fcakyon in https://github.com/ultralytics/ultralytics/pull/25100
• Fix silent FP32 fallback for legacy precision flags in benchmark by @JESUSROYETH in https://github.com/ultralytics/ultralytics/pull/25106
• Speed up segment2box by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/25095
• Fix NaN semantic loss when the whole batch is ignore pixels by @JESUSROYETH in https://github.com/ultralytics/ultralytics/pull/25097
• Validate bool-or-str CLI args (compile) at cfg layer (#25056) by @Joe-KI333 in https://github.com/ultralytics/ultralytics/pull/25082
• Refactor segmentation dataset documentation by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/25063
• Validate integer minimums and compile type at cfg layer by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25114
• Fail cleanly for optimize=True on PyTorch builds without XNNPACK by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25116
• Fall back to eager when torch.compile fails at first forward by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25115
• Fix ValueError when predict is called with an empty embed list by @ahmet-f-gumustas in https://github.com/ultralytics/ultralytics/pull/25123
• Fix segment mask conversion output directory by @zhanghuiwan in https://github.com/ultralytics/ultralytics/pull/25117
• Upgrade Python Docker image to 3.13 Trixie by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25127
• Add Apple Core AI integration guide by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25129
• Fix raise DeprecationWarning halting classify_transforms by @Dheeraj-Bhaskaruni in https://github.com/ultralytics/ultralytics/pull/24236
• Update ultralytics-inference version to 0.0.27 in documentation by @onuralpszr in https://github.com/ultralytics/ultralytics/pull/25130
• Preserve save_dir override when resuming training by @yegeniy in https://github.com/ultralytics/ultralytics/pull/23623
• Upsample before crop in process_mask to prevent mask leakage by @wjddnwp29 in https://github.com/ultralytics/ultralytics/pull/24463
• Accept common truthy strings for boolean environment variables by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/24730
• Document .txt image-list usage in dataset YAML splits by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/24472
• Reject undersized Axelera exports by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25131
• Fix YOLOE tensor input support with visual prompts by @emmanuel-ferdman in https://github.com/ultralytics/ultralytics/pull/22251
• Support optional kpt_oks_sigmas for pose evaluation via dataset YAML by @cosmo-gb in https://github.com/ultralytics/ultralytics/pull/24381
• Enable dynamic shapes and NMS for Alibaba MNN export by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/22348

New Contributors

• @Dheeraj-Bhaskaruni made their first contribution in https://github.com/ultralytics/ultralytics/pull/24236
• @yegeniy made their first contribution in https://github.com/ultralytics/ultralytics/pull/23623
• @Joe-KI333 made their first contribution in https://github.com/ultralytics/ultralytics/pull/25082
• @wjddnwp29 made their first contribution in https://github.com/ultralytics/ultralytics/pull/24463
• @cosmo-gb made their first contribution in https://github.com/ultralytics/ultralytics/pull/24381

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.92...v8.4.93

🌟 Summary

Ultralytics v8.4.92 improves RKNN INT8 multi-batch export reliability, strengthens inference and dataset validation, and expands YOLO26 documentation for web deployment. 🚀

📊 Key Changes

• 🔧 Fixed RKNN INT8 multi-batch export (PR #25094, @glenn-jocher)

  • RKNN calibration now uses a batch-1 ONNX graph, while the requested inference batch is passed to RKNN Toolkit through rknn_batch_size.
  • Resolves export failures when using configurations such as batch=8.
  • Adds regression coverage to verify that calibration and runtime batch handling remain correct.

• ⚡ Improved compiled model predictor reuse (PR #25092, @glenn-jocher)

  • Predictor initialization now explicitly checks whether the model is None.
  • Prevents torch.compile model wrappers from triggering unsupported truth-value checks during repeated inference.

• ✅ More accurate validation split handling (PR #25093, @glenn-jocher)

  • Dataset validation now checks the split requested by the user, such as val or test, instead of always checking val.
  • Missing or unavailable splits now produce a clear error rather than silently using another split.

• 📦 More reliable polygon-to-box conversion (PR #25086, @JESUSROYETH)

  • segment2box() now preserves the visible portion of polygons that cross image boundaries.
  • Improves bounding boxes after perspective and other augmentations, reducing shrunken or collapsed boxes.

• 🌐 Expanded LiteRT web deployment documentation (PR #25087, @onuralpszr)

  • Documents browser and Node.js inference with LiteRT.js and the official @ultralytics/yolo NPM package.
  • Includes WebGPU acceleration with CPU/WASM fallback and support for all six YOLO26 tasks.
  • Installation is documented with npm i @ultralytics/yolo @litertjs/core.

• 📚 Refreshed pose dataset documentation

  • Corrects dataset sizes, splits, keypoint details, and download requirements for COCO-Pose, COCO8-Pose, Dog-Pose, Hand Keypoints, and Tiger-Pose.
  • Adds clearer guidance for using Ultralytics Platform to annotate and train custom pose datasets.
  • Clarifies that COCO-Pose requires approximately 27 GB for its complete first download.

• 🔐 Simplified contributor license workflow

  • Replaces the third-party CLA action with the shared ultralytics/actions workflow.
  • Reduces duplicated configuration and limits workflow permissions.

🎯 Purpose & Impact

• Rockchip users can export INT8 models with larger inference batches more reliably, especially for edge deployments using RKNN hardware. 🧩
• Compiled models are safer to reuse across multiple predictions, improving stability for optimized inference workflows.
• Dataset validation errors are clearer and more trustworthy, helping users catch incorrect or missing test or val data earlier.
• Augmented segmentation data retains correct object extents, which can improve training labels and downstream detection quality near image borders.
• YOLO26 deployment is easier in browsers, enabling serverless webcam, image, and real-time inference through WebGPU or CPU/WASM fallback. 🌍
• Pose estimation documentation is more accurate and practical, making dataset selection, download planning, and training setup easier for both new and experienced users.

What's Changed

• Replace third-party CLA action by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25084
• docs: 📝 Update LiteRT documentation to include Web deployment details and npm package installation by @onuralpszr in https://github.com/ultralytics/ultralytics/pull/25087
• Fix compiled model predictor reuse by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25092
• Validate requested detection dataset split by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25093
• Fix segment2box shrinking boxes when augmented polygons cross image bounds by @JESUSROYETH in https://github.com/ultralytics/ultralytics/pull/25086
• Refactor pose dataset docs by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/25062
• Fix RKNN INT8 multi-batch export by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25094

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.91...v8.4.92

🌟 Summary

Ultralytics v8.4.91 improves YOLO26 deployment quality—especially TensorRT INT8 confidence calibration—while adding stronger automated CLI fuzz testing, better export reliability, and many training/inference bug fixes 🚀

📊 Key Changes

• Improved TensorRT INT8 exports for YOLO26 confidence scores ⚡

  • Excludes the Sigmoid operation from INT8 quantization in TensorRT export paths.
  • Applies higher precision for Sigmoid across TensorRT 7 through TensorRT 11+ using the appropriate backend mechanisms.
  • Adds GPU export test coverage for static INT8 detection exports.

• Added daily YOLO CLI fuzz testing 🧪

  • Introduces a new Monte Carlo fuzzing workflow for the yolo CLI across train, export, predict, validation, and chaos-style command variations.
  • Automatically classifies failures, confirms reproducible issues, deduplicates findings, and can file GitHub issues for confirmed bugs.

• Improved ONNX INT8 export reliability 🔧

  • ONNX INT8 quantization now focuses only on weighted operations such as Conv, Gemm, and MatMul.
  • INT8 ONNX exports are capped to a compatible opset to avoid unsupported ONNX Runtime static quantization paths.

• Better pretrained fine-tuning behavior 🎯

  • Adds class-name-based remapping for pretrained detection heads when fine-tuning on datasets with reordered or partially matching class names.
  • Helps transfer learned weights more accurately instead of relying only on class index order.

• Multiple inference, validation, and tracking fixes ✅

  • Fixes track() crashes caused by CPU/GPU tensor mismatches when using exported models with the default tracker.
  • Fixes classes filtering being applied after max_det truncation for end-to-end NMS models, which could return too few or zero detections.
  • Fixes pose validation save_txt=True keypoint scaling so saved keypoints correctly align with original image coordinates.
  • Fixes several Ultralytics Solutions crashes with OBB models, including Heatmap and ObjectBlurrer.

• Improved support for non-RGB and custom-channel workflows 🌈

  • Fixes knowledge distillation warmup for grayscale and other non-3-channel datasets.
  • Fixes TensorBoard graph logging for multispectral or custom-channel training.
  • Fixes classification transforms so custom interpolation is respected for non-square image sizes.

• Export and platform reliability improvements 🧩

  • Edge TPU compiler failures now raise clear errors instead of returning corrupt stub files.
  • ExecuTorch and ARM export dependencies are pinned more safely to avoid upstream Torch compatibility breakage.
  • Version comparison logic now handles 4-part package versions like 4.13.0.90 correctly.
  • Windows OpenVINO CI coverage is kept while limiting CPU instruction dispatch for more stable tests.

• Documentation refreshes 📚

  • Large updates to classification, detection, and OBB dataset docs with verified image counts, split details, clearer usage examples, and better dataset descriptions.
  • Adds or updates YOLO26 tutorial videos, including semantic segmentation documentation.
  • Clarifies Ultralytics Platform data residency details and improves Hailo docs formatting.

• CI and Docker workflow hardening 🐳

  • Codecov uploads are now best-effort, so external reporting outages do not fail otherwise successful test jobs.
  • Docker validation containers are auto-removed and test assets are pre-cached before running tests.

🎯 Purpose & Impact

• More accurate TensorRT INT8 confidence calibration 📈

  • Keeping Sigmoid at higher precision helps prevent confidence-score compression in YOLO26 TensorRT INT8 exports.
  • This can improve F1 calibration and reduce deployment surprises for users relying on INT8 acceleration.

• More reliable production exports 🚀

  • ONNX, TensorRT, Edge TPU, ExecuTorch, OpenVINO, and ARM-related fixes reduce export failures and make deployment behavior clearer across platforms.

• Fewer silent or confusing failures 🛡️

  • Invalid CLI/config values, failed Edge TPU compilers, unsupported dependency combinations, and broken tracking/device paths are now handled more robustly.

• Better results when fine-tuning custom datasets 🎓

  • Class-name-aware head remapping helps users reuse pretrained models more effectively when dataset class order differs from the original model.

• Stronger support for specialized vision data 🛰️

  • Grayscale, multispectral, OBB, pose, and custom-channel workflows receive important fixes, helping users working beyond standard RGB detection.

• Higher long-term stability through fuzzing 🧪

  • The new scheduled CLI fuzz workflow should uncover edge-case bugs that normal test matrices miss, improving future release quality for both developers and end users.

• Clearer learning and dataset guidance 📖

  • Refreshed docs make it easier to choose datasets, understand dataset structure, train YOLO26 models, and use Ultralytics Platform for annotation, training, and deployment.

What's Changed

• Use text fence for hailortcli device output in Hailo docs by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25041
• Fix track() device mismatch with exported models under the default tracktrack tracker by @JESUSROYETH in https://github.com/ultralytics/ultralytics/pull/25042
• Add https://youtu.be/_fvGA9LPXzs to docs by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/25048
• Fix classify_transforms ignoring interpolation for non-square sizes by @ahmet-f-gumustas in https://github.com/ultralytics/ultralytics/pull/25047
• Fix crashes on invalid copy_paste_mode and optimizer values by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25052
• Update data region warning with additional details by @sokrisba in https://github.com/ultralytics/ultralytics/pull/25050
• Fix classes filter applied after max_det truncation in end-to-end NMS by @JESUSROYETH in https://github.com/ultralytics/ultralytics/pull/25043
• Remap pretrained cls head by class name when fine-tuning by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/25051
• Quantize only weighted ops for ONNX INT8 by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/25046
• Fix Heatmap, ObjectBlurrer and other solutions crash with OBB models by @JESUSROYETH in https://github.com/ultralytics/ultralytics/pull/25058
• Fix save_txt keypoint scaling in PoseValidator by @JESUSROYETH in https://github.com/ultralytics/ultralytics/pull/25053
• Fix incorrect bin indexing in plt_color_scatter by @dorianboille in https://github.com/ultralytics/ultralytics/pull/25064
• Refactor OBB dataset documentation by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/25061
• Refactor classification dataset documentation by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/25003
• Refactor detect dataset documentation by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/25013
• Fix ExecuTorch ARM torch compatibility by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25069
• Reject None for typed CLI defaults by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25070
• Raise Edge TPU compiler failures instead of returning corrupt stub by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25068
• Bump onnx from 1.21.0 to 1.22.0 in /examples/RTDETR-ONNXRuntime-Python in the pip group across 1 directory by @dependabot[bot] in https://github.com/ultralytics/ultralytics/pull/25071
• Skip OpenVINO inference on Windows CI by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25072
• Keep Windows OpenVINO inference tests by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25073
• Make Codecov uploads best effort by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25075
• Fix DistillationModel warmup for single-channel datasets by @JESUSROYETH in https://github.com/ultralytics/ultralytics/pull/25067
• Pre-cache Docker test assets by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25074
• Fix Tensorboard logging with multi-spectral training by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/25076
• Fix version comparisons for 4-segment pins like opencv-python!=4.13.0.90 by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25079
• Pin ARM export torch below 2.13 by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25080
• Add Fuzz workflow: daily Monte Carlo fuzzing of the yolo CLI by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25049
• Handle v-prefixed versions in check_version by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25083
• Exclude Sigmoid/Softmax from TensorRT INT8 quantization by @davidnichols-ops in https://github.com/ultralytics/ultralytics/pull/25020

New Contributors

• @sokrisba made their first contribution in https://github.com/ultralytics/ultralytics/pull/25050
• @davidnichols-ops made their first contribution in https://github.com/ultralytics/ultralytics/pull/25020
• @dorianboille made their first contribution in https://github.com/ultralytics/ultralytics/pull/25064

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.90...v8.4.91

🌟 Summary

Ultralytics v8.4.90 adds a major new YOLO Architecture Explained guide for understanding the evolution from YOLOv3 to YOLO26, while also improving tracking reliability, segmentation edge-case handling, Docker GPU guidance, and contributor workflows 🚀📚

📊 Key Changes

• 🧠 New YOLO Architecture Explained guide

  • Added a detailed documentation page explaining how Ultralytics YOLO architectures evolved from YOLOv3 → YOLOv5 → YOLOv8 → YOLO11 → YOLO26.
  • Covers the major building blocks across the backbone, neck, and detection head, including Bottleneck, C3, C2f, C3k2, SPPF, C2PSA, DFL, anchor-free detection, and YOLO26’s NMS-free / DFL-free design.
  • Includes stage-level Mermaid diagrams, model inspection details, and links to the relevant YAML configs and module definitions.
  • Registered in mkdocs.yml and the guides index for easier discovery.
  • Contributed by @raimbekovm.

• 🎯 Improved ByteTrack and FastTrack low-confidence recovery

  • Fixed second-stage association in BYTETracker and FastTracker so low-confidence detections can correctly recover existing tracks when fuse_score=True.
  • The second association now uses IoU-only matching, matching ByteTrack’s intended behavior for low-score detections.
  • Added tests to confirm low-confidence detections keep the same track ID.

• 🧩 Segmentation mask utilities now handle zero detections

  • Fixed process_mask, process_mask_native, and scale_masks for valid empty inputs with zero detections.
  • These functions now return correctly shaped empty tensors instead of crashing.
  • Simplified segmentation validation postprocessing to rely on the shared utility behavior.

• 🐳 Docker GPU examples updated to CDI device requests

  • Replaced legacy --runtime=nvidia --gpus all examples with modern CDI-style device requests such as --device nvidia.com/gpu=all.
  • Updated Dockerfiles, quickstarts, Triton docs, DGX Spark docs, and YOLOv5 Docker documentation.
  • Added guidance that CDI requires Docker >= 28.2.0 and NVIDIA Container Toolkit >= 1.18, with legacy flags still noted for older systems.

• 🤖 New AI-agent contributor guidance

  • Added AGENTS.md with repository overview, engineering principles, development commands, PR workflow expectations, and architecture notes.
  • Added CLAUDE.md as a symlink for Claude Code compatibility.
  • Helps AI coding agents and contributors work more consistently in the Ultralytics repository.

• 🔇 Cleaner safe-load behavior

  • Removed an unnecessary warning when restricted model loading is unavailable on older PyTorch versions.
  • Reduces log noise without changing model loading behavior.

• 🧪 CI compatibility fix

  • Pinned NumPy <2 for the PyTorch 2.3.0 / torchvision 0.18.0 slow-test shard to avoid a known ColorJitter hue overflow issue.
  • Improves CI stability for legacy dependency combinations.

• 🔗 Documentation link updates and cleanup

  • Updated renamed Ultralytics solutions URLs from old ai-in-* paths to new computer-vision-in-* paths.
  • Cleaned up SolutionResults docs table formatting.
  • Simplified Roboflow-100 docs example by importing Roboflow directly and clarifying API key requirements.
  • Added contributor attribution for SuperMarioYL.

🎯 Purpose & Impact

• 📚 Easier learning and model understanding

  • The new architecture guide makes YOLO’s evolution easier to understand for both researchers and practical users.
  • Users can better see why YOLO26 is faster and simpler at deployment, especially with its end-to-end, NMS-free detection design.

• 🚀 More reliable object tracking

  • Tracking applications such as traffic analytics, sports analysis, retail monitoring, and security systems should better preserve object IDs when detections briefly drop in confidence.

• ✅ More robust segmentation workflows

  • Developers manually using mask post-processing utilities, especially with exported models, can now safely process frames where no objects are detected.

• 🐳 More stable GPU containers

  • CDI-based Docker examples help long-running GPU workloads, CI runners, and training containers keep GPU access more reliably across host system reloads.

• 🤝 Better contributor and automation support

  • AGENTS.md gives human contributors and AI coding tools clearer expectations, helping future PRs stay consistent, tested, and maintainable.

• 🧹 Smoother user experience

  • Reduced warnings, cleaner docs, updated links, and improved tests make the package easier to use, maintain, and troubleshoot.

What's Changed

• Silence safe load fallback warning by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25033
• Add AGENTS.md agent guidance with CLAUDE.md symlink by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25035
• Lead AGENTS.md with repository overview by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25036
• Fix ByteTrack second association ignoring low-confidence detections with fuse_score=True by @JESUSROYETH in https://github.com/ultralytics/ultralytics/pull/25034
• Fix process_mask, process_mask_native and scale_masks on empty 0-detection inputs by @JESUSROYETH in https://github.com/ultralytics/ultralytics/pull/25032
• fix(ci): pin NumPy <2 for torch 2.3.0 / torchvision 0.18.0 SlowTests shard by @onuralpszr in https://github.com/ultralytics/ultralytics/pull/25037
• docs: update links to renamed solution pages by @miles-deans-ultralytics in https://github.com/ultralytics/ultralytics/pull/25039
• Replace legacy --gpus examples with CDI device requests across docs and Dockerfiles by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25040
• Add YOLO architecture explained guide by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/24970

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.89...v8.4.90

🌟 Summary

Ultralytics v8.4.89 improves prediction reliability for grayscale NumPy images, adds several important edge-case fixes, and makes training, benchmarking, AI Gym, and release publishing more robust 🚀

📊 Key Changes

• 🖼️ Fixed 2D grayscale NumPy prediction on color models by @maxime2476

  • Raw grayscale NumPy arrays shaped like (H, W) are now expanded correctly for 3-channel color models.
  • This makes NumPy inputs behave consistently with PIL images and file paths.
  • Prevents PyTorch channel mismatch crashes during prediction.

• 🏋️ Fixed AI Gym result alignment by @SuperMarioYL

  • workout_count, workout_stage, and workout_angle now only report currently visible tracked people.
  • Results now stay aligned with total_tracks, avoiding stale data from people who already left the frame.

• ⚡ Improved multi-GPU training performance by @ExtReMLapin

  • Distributed training now uses broadcast_buffers=False by default.
  • This can reduce unnecessary GPU synchronization overhead, especially on non-NVLink systems or high-resolution training.

• 🧪 Improved benchmark usability by @zhanghuiwan and @raimbekovm

  • Benchmark format values are now case-insensitive, so inputs like ONNX or TensorRT work as expected.
  • Benchmark docs now include the eps argument and clarify standalone benchmark() defaults, including model="yolo26n.pt" and imgsz=160.

• 🌍 Fixed YOLOE and YOLO-World CLI class parsing by @ahmet-f-gumustas

  • Class names passed through the CLI now strip extra whitespace.
  • For example, classes="person, bus" now becomes ["person", "bus"] instead of ["person", " bus"].

• 🔗 Fixed signed model URL suffix validation by @diaz3z

  • URLs like model.pt?token=abc now pass .pt suffix checks correctly.
  • This improves compatibility with private or authenticated model downloads.

• 📐 Fixed segment2box() for objects on the left image edge by @bujna94

  • Segments where all x-coordinates are 0 are no longer incorrectly dropped.
  • This improves bounding box generation for objects touching the image border.

• 🧩 Improved installation reliability by @Nailujj

  • Excludes the known problematic opencv-python==4.13.0.90 package.
  • Helps avoid OpenCV-related crashes in environments such as Databricks.

• 🛠️ Hardened release publishing workflow by @glenn-jocher

  • Manual recovery runs are now safer after partial release failures.
  • Existing PyPI artifacts and release uploads are handled more gracefully.
  • Slack notifications now wait for publishing and SBOM generation to finish.

🎯 Purpose & Impact

• ✅ More reliable inference inputs
  Users can now pass grayscale NumPy arrays directly to standard color YOLO models without manual channel conversion.

• 🔄 More consistent behavior across input types
  NumPy, PIL, and file-based grayscale images are now normalized more consistently before prediction.

• 🚀 Better distributed training efficiency
  Multi-GPU users may see reduced synchronization bottlenecks, especially in bandwidth-limited setups.

• 📊 Clearer and more forgiving benchmarking
  Benchmark workflows are easier to use, with fewer avoidable errors from capitalization or undocumented defaults.

• 🏋️ Cleaner AI Gym analytics
  Workout tracking outputs are easier to consume because per-person lists now match the people currently visible in the frame.

• 🔐 Better support for real-world deployment workflows
  Signed URLs and managed Python environments are handled more reliably, reducing friction for cloud and enterprise users.

• 🧪 Stronger regression coverage
  New tests protect the grayscale NumPy fix, signed URL handling, segment2box() border behavior, and CLI class parsing from future regressions.

What's Changed

• Harden publish workflow recovery, notifications, and dispatch guard by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25027
• Add GhostNetV2 (DFC attention) modules and YOLOv8 GhostV2 configs by @kaanrkaraman in https://github.com/ultralytics/ultralytics/pull/23108
• Update README by @versa-dev in https://github.com/ultralytics/ultralytics/pull/23656
• Strip whitespace from CLI classes for YOLOE/World by @ahmet-f-gumustas in https://github.com/ultralytics/ultralytics/pull/24677
• Exclude broken opencv-python 4.13.0.90 wheel by @Nailujj in https://github.com/ultralytics/ultralytics/pull/23376
• Fix AIGym workout result lists to align with visible track IDs by @SuperMarioYL in https://github.com/ultralytics/ultralytics/pull/24228
• Strip URL query strings in check_suffix for signed model URLs by @diaz3z in https://github.com/ultralytics/ultralytics/pull/24369
• Set DDP broadcast_buffers=False by default by @ExtReMLapin in https://github.com/ultralytics/ultralytics/pull/24412
• Make benchmark format comparison case-insensitive by @zhanghuiwan in https://github.com/ultralytics/ultralytics/pull/24453
• Document benchmark eps arg and standalone benchmark() defaults by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/24560
• Fix segment2box dropping segments on the left image edge by @bujna94 in https://github.com/ultralytics/ultralytics/pull/24679
• Fix 2D grayscale NumPy array prediction on color models by @maxime2476 in https://github.com/ultralytics/ultralytics/pull/24751

New Contributors

• @SuperMarioYL made their first contribution in https://github.com/ultralytics/ultralytics/pull/24228
• @Nailujj made their first contribution in https://github.com/ultralytics/ultralytics/pull/23376
• @maxime2476 made their first contribution in https://github.com/ultralytics/ultralytics/pull/24751
• @versa-dev made their first contribution in https://github.com/ultralytics/ultralytics/pull/23656
• @diaz3z made their first contribution in https://github.com/ultralytics/ultralytics/pull/24369
• @zhanghuiwan made their first contribution in https://github.com/ultralytics/ultralytics/pull/24453
• @bujna94 made their first contribution in https://github.com/ultralytics/ultralytics/pull/24679

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.88...v8.4.89

🌟 Summary

Ultralytics v8.4.88 focuses on more reliable training for tiny datasets, stable NCNN export, and clearer documentation for model training concepts. 🚀

📊 Key Changes

• Tiny dataset dataloader fix by @glenn-jocher 🧠

  • Dataloader worker processes are now capped based on the actual number of final batches.
  • Single-batch training loaders now run in-process instead of spawning extra persistent workers.
  • Empty dataset handling now correctly relies on PyTorch DataLoader validation instead of triggering internal worker-count errors.
  • Added regression tests for tiny loaders, distributed sampling with drop_last, and empty dataset behavior.

• Improved NCNN export stability by @glenn-jocher 🛠️

  • PNNX is pinned to 20260526 for NCNN export and CI checks.
  • This avoids a known newer PNNX regression that could cause NCNN inference crashes.
  • NCNN remains paired with the latest available ncnn wheel.

• Clearer training documentation for dfl by @fcakyon 📚

  • The dfl training argument description now correctly explains Distribution Focal Loss as a bounding box localization term, not a classification feature.
  • This helps users better understand how YOLO models learn accurate box positions.

• Knowledge distillation guide visual update by @RizwanMunawar 🖼️

  • Added a new workflow image to the knowledge distillation documentation.
  • The image helps explain the teacher-student model training concept more intuitively.

• Version bump 📦

  • Package version updated from 8.4.87 to 8.4.88.

🎯 Purpose & Impact

• Better experience for small experiments and quick tests ⚡
  Users training on very small datasets should see fewer unnecessary background worker processes, lower overhead, and fewer cases where stuck workers keep CUDA resources busy.

• More robust distributed and edge-case training behavior ✅
  The dataloader update preserves existing drop_last behavior while improving worker handling after sampler construction, making training safer across normal, tiny, and distributed dataset setups.

• More dependable NCNN deployment workflows 📱
  Pinning PNNX improves reliability for users exporting YOLO models, including YOLO26 models, to NCNN for lightweight or edge-device inference.

• Clearer learning resources for all users 🌐
  Documentation updates make advanced concepts like knowledge distillation and Distribution Focal Loss easier to understand for both new users and experienced developers.

What's Changed

• Add cover image for knowledge-distillation.md by @RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/25026
• Pin PNNX for NCNN export stability by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25028
• Clarify DFL argument describes box localization not classification by @fcakyon in https://github.com/ultralytics/ultralytics/pull/25029
• Cap dataloader workers for tiny datasets by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25031

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.87...v8.4.88

🌟 Summary

🚀 Ultralytics v8.4.87 delivers a cleaner, safer GPU device-selection system, plus stability and performance fixes for training, inference, tracking, exports, and dataset checks.

📊 Key Changes

• Clean-sheet CUDA device selection 🧭

  • Added parse_device() to normalize device inputs such as cuda:0, 0,1, lists/tuples, torch.device, and -1 idle-GPU auto-selection.
  • select_device() no longer mutates CUDA_VISIBLE_DEVICES, making device selection predictable across repeated calls and long-running Python processes.
  • Explicit single-GPU requests now use torch.cuda.set_device() instead of environment-variable remapping.
  • Trainer, DDP setup, validation, autobatch, and distributed barriers now consistently use resolved CUDA device indices.
  • Added documentation for ultralytics.utils.torch_utils.parse_device.

• Stronger GPU training tests 🧪

  • Added a cold-process nonzero-GPU training test to better match real CLI and Ultralytics Platform training behavior.
  • Verifies that training on GPUs like device=1 or higher works correctly from a fresh process without relying on previous CUDA initialization.

• Fixed DataLoader worker cleanup at training shutdown 🧹

  • Added a close() method to InfiniteDataLoader.
  • Training now explicitly shuts down persistent train and validation workers before Python exits.
  • Helps prevent end-of-run DataLoader worker ... killed by signal: Terminated errors after results are already saved.

• Improved inference warmup for standard NMS ⚡

  • AutoBackend.warmup() now preloads torchvision for non-end-to-end models.
  • This helps later non-max suppression calls use faster torchvision NMS when appropriate, reducing first-inference latency after warmup.

• Corrected dataset file-speed reporting 💾

  • Fixed an inverted condition in check_file_speeds().
  • Slow storage, such as network-mounted datasets, should now trigger the intended warning instead of being incorrectly reported as “Fast image access ✅”.

• Tracking ReID device alignment 🎯

  • Trackers now pass the predictor device into ReID encoders.
  • ReID models are initialized and run on the same device as prediction where applicable, improving consistency for tracking workflows.

• Export reliability improvements 📦

  • TensorFlow SavedModel export now distinguishes CUDA vs non-CUDA export paths more carefully.
  • CPU exports hide TensorFlow GPUs where possible to avoid unnecessary GPU memory use.
  • ONNX Runtime and Paddle dependency checks now better handle interchangeable CPU/GPU package variants to avoid unnecessary or conflicting installs.
  • Paddle export now uses the actual export device to decide whether GPU Paddle is needed.

🎯 Purpose & Impact

• More reliable GPU behavior 🚀

  • Users should see fewer surprises when training, validating, predicting, or exporting repeatedly in the same Python session.
  • This is especially important for notebooks, services, CI, distributed training, and production systems where changing CUDA_VISIBLE_DEVICES mid-process can cause hard-to-debug issues.

• Better support for nonzero GPU training 🖥️

  • Training on GPUs beyond CUDA:0 is now more robust, including cold-start CLI usage common in production and Ultralytics Platform environments.

• Cleaner shutdowns after training ✅

  • Persistent DataLoader workers are now cleaned up explicitly, reducing noisy shutdown crashes and improving confidence that completed runs exit cleanly.

• Lower latency after warmup ⚡

  • Standard detection workflows can benefit from smoother post-warmup inference performance by ensuring faster NMS paths are ready when needed.

• More accurate dataset diagnostics 📊

  • Users with slow disks or network storage will receive correct warnings, helping them identify dataset I/O bottlenecks that can slow training.

• More consistent tracking and export workflows 🔄

  • ReID tracking components now better follow the selected prediction device.
  • Export paths are less likely to allocate unwanted GPU memory or install conflicting runtime packages.

What's Changed

• Add cold-process nonzero-device GPU train test by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25019
• Fix inverted read-speed condition in dataset file speed check by @ahmet-f-gumustas in https://github.com/ultralytics/ultralytics/pull/25025
• Fix leaked dataloader workers at end of training (atexit killed by signal: Terminated crash) by @Bovey0809 in https://github.com/ultralytics/ultralytics/pull/25024
• Preload torchvision during warmup for non-end2end NMS path by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/25023
• Clean-sheet device selection: stop mutating CUDA_VISIBLE_DEVICES by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25021

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.86...v8.4.87

🌟 Summary

Ultralytics v8.4.86 fixes a CUDA device selection regression that could interrupt training when using nonzero GPU indices, restoring reliable multi-GPU workflows. 🚀

📊 Key Changes

• Fixed CUDA device re-selection after remapping 🛠️
  Resolved an issue where training with a command like device=3 could fail later during validation or final evaluation after CUDA_VISIBLE_DEVICES remapped that GPU to cuda:0.

• Restored expected behavior for remapped GPUs 🔁
  If CUDA_VISIBLE_DEVICES already matches the requested device, Ultralytics now correctly recognizes that CUDA has already applied the remap and returns the proper visible device index.

• Added regression test coverage ✅
  A new test covers the specific case where a nonzero physical GPU, such as GPU 3, is remapped to the single visible CUDA device cuda:0 after CUDA initialization.

• Version bump 📦
  Updated the package version from 8.4.85 to 8.4.86.

🎯 Purpose & Impact

• Prevents unexpected training crashes 💥➡️✅
  Users training on a specific nonzero GPU, for example device=3, should no longer see invalid CUDA device errors during later training stages such as trainer.final_eval().

• Improves production reliability 🏭
  This is especially important for servers, clusters, and shared GPU environments where jobs are often assigned to GPUs using CUDA_VISIBLE_DEVICES.

• Keeps recent validation improvements intact 🔒
  The fix preserves the stricter device validation introduced previously while correcting the special case where CUDA remapping has already taken effect.

• Better support for multi-GPU systems 🖥️
  Developers and teams using machines with multiple GPUs can expect more predictable behavior when selecting specific GPU indices.

• No user action required beyond upgrading ⬆️
  Install the latest release with:

  pip install -U ultralytics

What's Changed

• Fix select_device re-request after CUDA_VISIBLE_DEVICES remap applied by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25018

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.85...v8.4.86

🌟 Summary

Ultralytics v8.4.85 focuses on reliability and deployment polish, with important fixes for training out-of-memory recovery, AutoBatch memory estimation, LiteRT INT8 calibration, dataset validation, YOLO-World validation, RT-DETR/CoreML behavior, and edge-device documentation 🚀

📊 Key Changes

• Fixed TaskAlignedAssigner OOM recovery during training 🧠
  The CPU fallback for TaskAlignedAssigner now runs after the failed CUDA traceback is released, preventing the recovery path itself from running out of GPU memory.

• Improved AutoBatch memory estimation 📦
  AutoBatch now better estimates detection loss memory usage, including the large temporary tensors used during assignment and classification loss calculation. This should help choose safer batch sizes for heavy training jobs.

• Improved LiteRT static INT8 calibration 📱
  LiteRT calibration now handles smaller or partial calibration batches by repeating samples to match the traced batch size instead of failing when the dataset does not provide a perfectly sized batch.

• Stricter dataset nc validation ✅
  Dataset YAML files now validate that nc is an integer or integer-like value. Invalid values such as placeholders or decimals fail early with a clearer error.

• YOLO-World standalone validation fix 🌍
  Added a dedicated WorldValidator so model.val() correctly loads dataset class names and regenerates text embeddings when validating YOLO-World models on custom datasets.

• Rectangular training shape fix 📐
  Augmentations now respect rect=True batch shapes instead of incorrectly falling back to square sizing, improving training consistency for datasets with varied aspect ratios.

• RT-DETR validation and docs cleanup ⚡
  RT-DETR validation transforms were simplified by removing a special stretch path, and docs now clarify that max_det cannot make pretrained RT-DETR models return more detections than their decoder query limit.

• CoreML RT-DETR inference improvement 🍎
  CoreML export metadata now records the model head type, allowing RT-DETR models to use ComputeUnit.ALL. This improves RT-DETR CoreML speed and helps preserve FP16 accuracy compared with Neural Engine-only execution.

• CUDA device selection fix 🖥️
  select_device() now preserves the requested CUDA index after CUDA has already been initialized, avoiding confusing device remapping behavior.

• Tracking index bug fix 🎯
  ByteTrack now preserves original detection indices for low-confidence detections matched during second association, reducing callback/index mismatches in tracking workflows.

• Better user-facing prediction windows 🪟
  show=True display windows now auto-scale oversized images on Linux and Windows to avoid cropped previews.

• YOLOE visual prompting verbosity fix 🔇
  YOLOE visual prompting now respects verbose=False, preventing unwanted setup logs when users request quiet prediction.

• Dataset name convenience 🗂️
  Detection-style validation now supports bare dataset names like coco8 by resolving them to coco8.yaml when available.

• Training fails fast when labels are missing 🚨
  Training now raises an error when no labels are found, helping users catch dataset setup issues immediately. Unlabeled validation splits can still warn instead of failing.

• Updated edge deployment documentation 📚
  Hailo docs now include a cleaner YOLO11-to-HEF workflow, metadata.yaml handling, and expanded inference examples for images, videos, webcams, and Raspberry Pi cameras. LiteRT, TFLite, and QNN docs were refreshed with current YOLO26n Android benchmark context.

• Security warning for HTTP console log streaming 🔐
  ConsoleLogger now warns when logs are sent to plaintext http:// destinations, helping users avoid transmitting sensitive console output unencrypted.

🎯 Purpose & Impact

• More stable training on large or dense datasets 💪
  The OOM recovery and AutoBatch fixes directly target real training failures, making Ultralytics more robust for workloads with many objects, large batches, or high memory pressure.

• Fewer confusing export and deployment failures 📲
  LiteRT calibration is more forgiving, CoreML RT-DETR inference is more accurate and faster, and Hailo deployment docs are clearer about required files and metadata.

• Better validation across model families 🧪
  YOLO-World, RT-DETR, and detection-style datasets now behave more consistently, especially when users validate custom datasets or use simplified dataset names.

• Clearer errors for dataset problems 🛠️
  Missing labels and invalid nc values are caught earlier, reducing wasted training time and making configuration issues easier to fix.

• Improved day-to-day usability ✨
  Rectangular training, CUDA device selection, tracking callbacks, prediction display windows, and YOLOE quiet mode all receive practical fixes that make common workflows smoother.

• Stronger mobile and edge guidance 🚀
  Refreshed YOLO26n LiteRT/TFLite/QNN benchmark documentation and expanded Hailo examples help users make better deployment decisions across Android, Raspberry Pi, and Hailo hardware.

What's Changed

• Improve Hailo documentation and add more inference options by @lakshanthad in https://github.com/ultralytics/ultralytics/pull/24645
• Fix rect=True training producing square instead of rectangular tensors by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/25009
• Clarify RT-DETR max_det and query limit by @oaslananka in https://github.com/ultralytics/ultralytics/pull/24994
• Simplify RTDETR dataset transforms and remove stretch argument from v8_transforms by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/25011
• Clarify square image size requirement in benchmark docs by @oaslananka in https://github.com/ultralytics/ultralytics/pull/24993
• Add semantic mask boundary quality guidance by @oaslananka in https://github.com/ultralytics/ultralytics/pull/24976
• Scale window to prevent cropped preview on WIndows by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/24973
• Add WorldValidator for YOLO-World by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/24974
• Use CoreML ComputeUnit.ALL for RT-DETR inference by @artest08 in https://github.com/ultralytics/ultralytics/pull/24999
• Fix verbose bug YOLOE visual prompting by @YoannSME in https://github.com/ultralytics/ultralytics/pull/24814
• Raise exception when labels are missing by @alhatemmohanad-ui in https://github.com/ultralytics/ultralytics/pull/24760
• Skip unstable RTDETR CPU train on JetPack 5 by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24945
• Support dataset names without .yaml extension in train and val by @lakshanthad in https://github.com/ultralytics/ultralytics/pull/24765
• Fix initialized CUDA device selection by @oaslananka in https://github.com/ultralytics/ultralytics/pull/24987
• Warn on plaintext HTTP console log streaming destinations by @amanharshx in https://github.com/ultralytics/ultralytics/pull/25001
• Optimize the index for tracking target feedback by @zhixuwei in https://github.com/ultralytics/ultralytics/pull/24971
• Update LiteRT and TFLite benchmark docs by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25015
• Fix TAL OOM recovery, AutoBatch loss estimation, LiteRT INT8 calibration, and nc validation by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25016

New Contributors

• @YoannSME made their first contribution in https://github.com/ultralytics/ultralytics/pull/24814
• @alhatemmohanad-ui made their first contribution in https://github.com/ultralytics/ultralytics/pull/24760

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.84...v8.4.85

🌟 Summary

Ultralytics v8.4.84 improves training reliability, OBB stability, RT-DETR benchmarking accuracy, and LiteRT/mobile deployment documentation, with the most important fix ensuring valid training runs still save checkpoints even after transient NaN/Inf events 🚀

📊 Key Changes

• More reliable checkpoint saving during training 🛡️
  Fixed a production issue where training could complete successfully but fail at the end because no best.pt or last.pt checkpoint was saved. This happened when transient NaN/Inf values contaminated the EMA model state.

• Cleaner imgsz validation errors 🧭
  Invalid image size strings like imgsz=640x480 now raise a clear ValueError explaining the correct formats, instead of exposing a raw Python parsing error.

• Improved OBB training stability for tiny rotated boxes 📦
  Added safer handling in rotated bounding box loss calculations by applying a small floor inside probabilistic IoU covariance calculations. This helps prevent loss divergence when training on very small oriented objects.

• Fixed RT-DETR benchmark validation after export 📈
  Exported RT-DETR models are now reloaded with the correct RTDETR wrapper during benchmarking instead of YOLO, ensuring the proper validator and postprocessing are used. This fixes cases where benchmark mAP incorrectly showed 0.

• Axelera YOLO11 segmentation example improvements 🎥
  The examples/YOLO-Axelera-Python/yolo11-seg.py example now handles numeric camera sources correctly by mapping them to /dev/video<N> and using the OpenCV backend. Single-image sources also keep the display window open properly.

• LiteRT and mobile deployment docs expanded 📱
  Documentation now highlights the official Ultralytics YOLO Flutter plugin for running LiteRT .tflite exports on Android, including real-time camera inference, single-image prediction, GPU acceleration, and YOLO26 task support.

• New LiteRT performance guidance ⚡
  Added measured Android LiteRT performance tables for YOLO26n models across detection, segmentation, semantic segmentation, classification, pose, and OBB tasks, helping users understand real-world CPU/GPU mobile performance.

• TensorFlow.js and TF SavedModel docs updated 🔗
  Deprecated TensorFlow.js export guidance now points to Google’s official LiteRT.js web runtime documentation, and TF SavedModel docs consistently refer to LiteRT instead of TensorFlow Lite where appropriate.

• Documentation tooling maintenance 🧰
  Updated the MkDocs Ultralytics plugin requirement to >=0.2.5 and added missing GitHub author metadata.

🎯 Purpose & Impact

• Fewer failed training jobs ✅
  Users are less likely to lose completed training runs due to checkpoint-saving failures. Even if a temporary numerical issue appears in EMA/model tensors, Ultralytics now sanitizes the saved checkpoint instead of ending with no persistent weights.

• Better user experience for configuration mistakes 💬
  Clearer imgsz errors make it easier for both beginners and advanced users to fix incorrect command-line or Python arguments quickly.

• More stable OBB model training 🧱
  Datasets with tiny rotated objects should train more reliably, reducing the chance of unstable gradients or diverging losses in oriented bounding box workflows.

• Trustworthy RT-DETR benchmark results 🎯
  RT-DETR users should now see meaningful benchmark metrics after export instead of misleading zero-mAP results caused by the wrong validation pipeline.

• Smoother edge and mobile deployment path 📲
  The LiteRT documentation improvements make it easier to choose the right deployment route for Android, browser, Apple, and Qualcomm NPU targets, especially with YOLO26 as the recommended model family.

• Improved examples and docs reduce friction 🙌
  Axelera users get a more practical segmentation demo, while documentation updates help deployment users avoid dead links, outdated terminology, and unclear driver-version issues.

What's Changed

• Fix --source camera handling and single-image display in yolo11-seg example by @tatsuya-toyoda in https://github.com/ultralytics/ultralytics/pull/24969
• fix: RTDETR benchmark mAP=0 in benchmark by @artest08 in https://github.com/ultralytics/ultralytics/pull/24998
• Bump MkDocs plugin requirement by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25002
• Fix LiteRT.js documentation link by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25000
• Add mobile-apps notice to LiteRT docs and finish TF SavedModel LiteRT rename by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25004
• Fix sub-stride OBB loss divergence via probiou floor by @lmycross in https://github.com/ultralytics/ultralytics/pull/25007
• Fix EMA-NaN no-checkpoint loss and unhelpful imgsz error by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25005

New Contributors

• @tatsuya-toyoda made their first contribution in https://github.com/ultralytics/ultralytics/pull/24969

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.83...v8.4.84

🌟 Summary

🚀 Ultralytics v8.4.83 is led by a major export upgrade: the new Google LiteRT format now replaces the older separate TFLite and TF.js export paths, while several reliability and performance fixes make training, segmentation, and deployment more stable.

📊 Key Changes

• 🔄 New LiteRT export replaces legacy TFLite and TF.js workflows

  • Added a new format="litert" export option.
  • LiteRT becomes the unified path for mobile, edge, embedded, and browser deployment.
  • Legacy tflite and tfjs names still work, but now show deprecation warnings and redirect to LiteRT.
  • Added a dedicated LiteRT backend so exported .tflite models can be loaded for predict and val inside Ultralytics.

• 📱 Broader on-device deployment support

  • LiteRT uses Google’s newer runtime direction for .tflite models.
  • Supports multiple quantization modes, including:
    • standard INT8
    • mixed INT8 + 16-bit activations
    • dynamic INT8 without calibration data
  • This makes exports more flexible for mobile and edge hardware.

• 🌐 Browser deployment is simplified

  • Instead of needing a separate TF.js export, browser use is now handled through LiteRT.js with the same .tflite model.
  • This reduces duplication and makes web deployment easier to understand.

• 🧠 Segmentation memory use and speed improved

  • Segmentation mask post-processing was optimized to reduce peak memory use and speed up execution.
  • This helps avoid out-of-memory failures on large images or high-object-count scenes.

• 🛡️ Training checkpoint reliability improved

  • Fixed a long-standing issue where a bad EMA state could prevent any checkpoint from being saved.
  • Training runs are now less likely to fail at the finish line because of temporary numerical issues.

• ⚡ Mixed-precision attention made safer

  • Attention computations were reordered to reduce FP16 overflow risk.
  • This helps prevent inf/nan problems during AMP or half-precision training.

• 🧹 Multi-dataset and dataloader memory/file-handle fixes

  • MultiTrainer now avoids retaining large trainer objects unnecessarily.
  • DataLoader reset now properly shuts down old workers, preventing “too many open files” errors in long runs.

• 🎯 Classification INT8 calibration improved

  • Classification exports now use a classification-specific dataset pipeline for INT8 calibration rather than detection-style loading.
  • This should produce more correct and robust classification quantized exports.

• ✅ Better dataset validation

  • Segment tasks now fail early if users accidentally provide box-only labels instead of segmentation labels.
  • This prevents confusing downstream errors later in validation or training.

• 🌍 YOLO-World distributed training fix

  • Class setup is now applied on all DDP ranks, improving consistency in multi-GPU training.

🎯 Purpose & Impact

• For mobile, edge, and web developers 📦
  LiteRT is the biggest change in this release. It simplifies deployment by replacing two older export formats with one modern, unified export flow. That means less confusion, fewer export branches, and a cleaner path from training to deployment.

• For teams shipping browser apps 🌐
  You no longer need to think of browser support as a separate TF.js pipeline. A single LiteRT .tflite model can now cover both on-device and web use cases more cleanly.

• For users exporting quantized models ⚙️
  The added LiteRT quantization options offer more choices for balancing size, speed, and accuracy, especially on constrained hardware.

• For training stability 🛠️
  Several fixes reduce frustrating failures:

  • fewer broken checkpoints
  • fewer FP16 numerical issues
  • fewer leaked workers and file descriptors
  • more reliable multi-dataset training

• For segmentation users 🎭
  Lower memory usage and faster mask processing can make segmentation inference more practical on larger inputs and reduce OOM crashes.

• For dataset preparation 🧾
  Earlier validation catches common mistakes sooner, which saves debugging time and prevents invalid runs.

• Overall impact ✨
  This release is especially valuable for anyone deploying YOLO26 to mobile, embedded, edge, or browser environments, while also improving everyday robustness for training and segmentation workloads.

What's Changed

• Linkify PR numbers in publish Slack notifications by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24966
• Use ClassificationDataset for classify INT8 export calibration by @onuralpszr in https://github.com/ultralytics/ultralytics/pull/24972
• Add Trendshift badge to README files by @onuralpszr in https://github.com/ultralytics/ultralytics/pull/24979
• Set YOLO-World eval classes on all DDP ranks by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/24980
• Fix MultiTrainer memory retention by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24977
• Fix sticky-NaN EMA checkpoint loss and legacy safe-load rejection by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24986
• Reduce segmentation mask memory and speed (chunked upsample + separable crop_mask) by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24988
• Pre-scale attention query to prevent fp16 overflow by @fcakyon in https://github.com/ultralytics/ultralytics/pull/24990
• Shut down stale DataLoader workers on reset by @fcakyon in https://github.com/ultralytics/ultralytics/pull/24989
• Reject box-only labels for segment datasets by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24992
• Add C++ inference guidance to export docs by @oaslananka in https://github.com/ultralytics/ultralytics/pull/24982
• Add tracker state guidance to tracking docs by @oaslananka in https://github.com/ultralytics/ultralytics/pull/24984
• Guard fuse_deconv_and_bn against non-BatchNorm bn by @fcakyon in https://github.com/ultralytics/ultralytics/pull/24991
• New Google LiteRT export replaces legacy TFLite and TF.js by @ambitious-octopus in https://github.com/ultralytics/ultralytics/pull/22870

New Contributors

• @oaslananka made their first contribution in https://github.com/ultralytics/ultralytics/pull/24982

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.82...v8.4.83

🌟 Summary

🔧 v8.4.82 is mainly a stability and usability release focused on fixing YOLO26 Axelera export issues, improving classification RAM caching, and tightening several training/data handling edge cases.

📊 Key Changes

• 🚀 Major fix for YOLO26 Axelera exports

  • Fixed a bug where Axelera export could accidentally delete its own output files when run from the model’s directory.
  • This was the key issue behind failed exports showing errors like “output model too small” even when compilation had actually succeeded.
  • The fix now uses a safer temporary compile folder and adds export serialization to avoid conflicts from shared process state.

• 🔄 Axelera export behavior restored for end2end=False

  • A temporary restriction added in the previous change was removed.
  • YOLO26 Axelera exports can again work with end2end=False where supported, avoiding unnecessary breakage in existing workflows such as Ultralytics Platform jobs.

• 🧠 Classification cache='ram' re-enabled with a memory-safe design

  • Classification training can now use RAM caching again without the previous memory growth problem.
  • Images are decoded once into a shared memory buffer instead of being copied separately into each worker.

• 🛑 Training now fails fast when all labels are empty

  • If a detection dataset has no usable labels at all, training now stops immediately with a clear error instead of just warning and failing later in confusing ways.

• 🖼️ Image format support cleanup

  • Added proper fallback loading for .heif images.
  • Removed the invalid .jpeg2000 extension from supported image format lists.

• 🧹 Python reliability improvements

  • Fixed several mutable default argument definitions across metrics, TAL, augmentations, YOLOE, and SAM-related code.
  • These bugs were subtle, but could cause shared state to leak across calls in long-running sessions.

• 📚 Docs and workflow polish

  • Clarified stream=True behavior in predict mode.
  • Documented MLFLOW_KEEP_RUN_ACTIVE for MLflow users.
  • Updated citations for YOLOv7, YOLOv9, YOLOv10, and YOLO12 docs.
  • Simplified Slack release/CI notifications for cleaner team alerts.

🎯 Purpose & Impact

• ✅ More reliable Axelera exports for YOLO26

  • This is the most important change in the release.
  • Users exporting YOLO26 models for Axelera hardware should see far fewer mysterious export failures, especially in automated or isolated environments like the Ultralytics Platform.

• ⚙️ Safer production workflows

  • The Axelera fix prevents exported models from being silently removed during cleanup.
  • That means more dependable deployment pipelines and less debugging time.

• 💾 Better memory efficiency for classification training

  • Re-enabling RAM caching can improve training speed while avoiding the old worker duplication problem.
  • This is especially helpful for users training classification models on larger datasets.

• 🧪 Earlier and clearer failure signals

  • Catching all-empty labels at the start helps users fix dataset problems right away instead of encountering vague downstream errors later.

• 📷 Smoother handling of real-world image files

  • Better HEIF support means fewer surprises when working with photos from modern devices.

• 🛠️ Improved long-running stability

  • The default-argument fixes reduce the chance of strange behavior caused by shared internal state, which matters most in notebooks, services, and repeated training/inference sessions.

Overall, v8.4.82 is less about new models and more about making existing YOLO26 export and training workflows more dependable 📦✨

What's Changed

• Guard YOLO26 Axelera end2end exports by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24962
• Restore Axelera end2end=false exports by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24963
• Fix mutable default args in metrics and tal by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/24631
• Fix mutable default args in augment, YOLOE, and SAM by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/24630
• Restore fail-fast when training labels are all empty by @AffanBinFaisal in https://github.com/ultralytics/ultralytics/pull/24562
• Docs: document MLFLOW_KEEP_RUN_ACTIVE env var for MLflow integration by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/24602
• docs: update YOLOv7, YOLOv9, YOLOv10 and YOLO12 citations by @banu4prasad in https://github.com/ultralytics/ultralytics/pull/24501
• docs: clarify stream=True behavior by @seungcle in https://github.com/ultralytics/ultralytics/pull/24657
• Drop phantom jpeg2000 extension and fix heif loader fallback by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/24594
• Refactor comments in tests/init.py for clarity by @penampungdipek-creator in https://github.com/ultralytics/ultralytics/pull/24651
• Fix classification cache='ram' leak via shared memory tensor by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/24670
• Simplify Slack notification messages by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24964
• Fix Axelera export deleting its output when run from the model directory by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24965

New Contributors

• @seungcle made their first contribution in https://github.com/ultralytics/ultralytics/pull/24657
• @penampungdipek-creator made their first contribution in https://github.com/ultralytics/ultralytics/pull/24651

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.81...v8.4.82

🌟 Summary

This release is mainly a stability and export-quality update for Ultralytics, led by an important FastSAM fix for multiple text prompts 🧠, plus clearer and stricter model export precision handling, better YOLO26 coverage, and several reliability improvements across platforms and integrations 🚀

📊 Key Changes

• 🛠️ FastSAM multiple text prompts no longer crash

  • Fixed a ValueError in FastSAMPredictor.prompt() when users passed more than one text prompt.
  • This especially affected common cases where some small masks were filtered out before prompt matching.
  • The fix vectorizes the index remapping, so lists like ["a photo of a dog", "a photo of a person"] now work as documented.

• 📦 Export precision behavior is now clearer and stricter

  • Export logic was cleaned up to use quantize= consistently instead of relying on older half=True / int8=True internal state.
  • Unsupported precision requests now fail clearly instead of silently changing to something else.
  • Examples:
    • CoreML FP32 stays FP32 when requested
    • CPU TorchScript FP16 now raises a clear error instead of quietly falling back
    • Format-specific precision support is better documented

• 📘 Export docs got a major accuracy upgrade

  • Many integration docs were updated to explain exactly which precisions each format supports, including:
    • CoreML
    • TorchScript
    • OpenVINO
    • MNN
    • RKNN
    • TensorFlow formats
    • ExecuTorch
    • Sony IMX500
  • This makes deployment choices easier and reduces trial-and-error.

• 🤖 YOLO26 became the main export smoke-test model in more places

  • Export smoke tests were updated from YOLO11 to YOLO26 for several formats.
  • This helps ensure the latest recommended model family is tested more directly in CI and export workflows.

• 🧬 Embedding behavior is more intuitive

  • model.embed() and model.predict() now behave better when used back-to-back.
  • After extracting embeddings, a normal model.predict() call returns standard Results again instead of unexpectedly staying in embedding mode.

• ⚡ Faster, safer imports for standard YOLO use

  • SAM/SAM3-related imports were made lazier, so normal Ultralytics imports won’t eagerly load optional torchvision pieces.
  • This fixes startup issues like missing torchvision::nms in environments that don’t need SAM at all.

• 🪟 Windows ONNXRuntime C++ example handles non-ASCII paths better

  • The ONNXRuntime C++ model loader now uses a Windows-safe filesystem path conversion.
  • This improves compatibility for model files stored in folders with UTF-8 or non-English characters.

• 🧩 Axelera export reliability improved

  • Added a numpy<=2.3.5 pin to avoid broken YOLO26 attention-graph exports.
  • Export output validation was tightened so suspiciously tiny models fail loudly instead of being reported as successful.

• 📄 Hailo documentation was clarified

  • The docs now more clearly explain that Hailo HEF is currently a manual ONNX-to-HEF workflow, not yet a direct model.export(format="hailo") target.
  • Notes were expanded around YOLO11, YOLO26, supported tasks, and Hailo hardware/SDK differences.

🎯 Purpose & Impact

• ✅ FastSAM is more usable for natural-language prompting

  • Users working with multiple text prompts can now rely on FastSAM without random crashes.
  • This is the most user-visible fix in the release and especially helpful for prompt-based segmentation workflows.

• 🔒 Exports are more trustworthy

  • If you request a precision the target format cannot actually support, Ultralytics now tells you immediately.
  • This prevents hidden precision mismatches that could lead to confusing model size, speed, or accuracy results.

• 🚀 YOLO26 deployment confidence improves

  • Since YOLO26 is the latest stable and recommended Ultralytics model, broader smoke-test coverage means better confidence when exporting it to different runtimes and hardware.

• 😊 Prediction and embedding workflows feel less surprising

  • Developers switching between feature extraction and normal inference should hit fewer gotchas.

• 🧯 Fewer environment-related failures

  • Lazy SAM imports reduce import-time breakage in setups that only need standard YOLO tasks.
  • Windows path fixes and stricter export validation also reduce frustrating edge-case failures.

• 📚 Better guidance for non-experts and deployers

  • The docs now do a better job of explaining what each export target really supports, helping users choose the right deployment format without guesswork.

Overall, v8.4.81 is a polish-and-reliability release ✨: it fixes an important FastSAM bug, makes export precision behavior much safer, and strengthens YOLO26 deployment readiness across the stack.

What's Changed

• Fix export precision handling and docs by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24950
• Fix Sentry bugs 2026-06-27 by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24949
• Remove internal half int8 export state by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24951
• Fix ONNXRuntime C++ model path on Windows by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24953
• Avoid bare alls variable in Hailo docs by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24954
• Update Hailo docs compatibility notes by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24955
• Pin Axelera numpy and remove redundant export tests by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24947
• Update export smoke models to YOLO26 by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24957
• Fix embed state reset for predictions by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24958
• Fix YOLO26 Axelera export output path by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24956
• Reduce slow CI Jetson TensorRT coverage by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24959
• Fix ValueError in FastSAM prompt() with multiple text prompts by @ahmet-f-gumustas in https://github.com/ultralytics/ultralytics/pull/24952

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.80...v8.4.81

🌟 Summary

🚀 Ultralytics v8.4.80 is mainly about making model export precision much simpler and more future-ready, with a new unified quantize argument replacing the older half and int8 switches—plus a few important reliability, validation, and training stability fixes.

📊 Key Changes

• New unified quantize export argument by @onuralpszr ⭐

  • Replaces the older half=True and int8=True style with one cleaner setting.
  • Supports simple values like:
    • quantize=16 for FP16
    • quantize=8 for INT8
    • quantize=32 for FP32
  • Also supports more advanced mixed-precision schemes like quantize="w8a16" where supported.
  • Older half and int8 arguments still work for now, so existing scripts should not break immediately. 🔄

• Export docs and integrations updated across the board

  • Many export examples and docs were refreshed to use quantize instead of half/int8.
  • This affects formats and platforms such as ONNX, TensorRT, OpenVINO, CoreML, TFLite, TF.js, RKNN, QNN, Axelera, DEEPX, and more. 📦

• Better support for modern and future quantization workflows

  • The new system is designed to scale beyond just “FP16 or INT8”.
  • This is especially useful as more hardware targets adopt mixed precision and platform-specific quantization modes. ⚙️

• Distributed validation now combines confusion matrices across GPUs

  • In multi-GPU validation, confusion matrix plots now reflect results from all workers, not just one process. 📈

• OBB training stability improved

  • A fix prevents some very tiny oriented bounding boxes from causing unstable or invalid loss behavior, including NaNs during training. 🛡️

• HUB/network request reliability improved

  • Added real request timeouts and retry behavior so stalled network calls are less likely to hang indefinitely. 🌐

• MLflow integration tests restored and modernized

  • Improves confidence that MLflow logging behavior continues to work correctly with newer MLflow versions. 🧪

• Tracking docs clarified

  • Better explanation of when to use persist=True and how to choose tracker backends like botsort.yaml or bytetrack.yaml. 🎯

• Ultralytics Platform dataset upload docs improved

  • Clearer explanation that browser video uploads depend on codec support, not just file extension.
  • AVI was removed from supported browser-upload guidance. 🎥

🎯 Purpose & Impact

• Simpler exporting for everyone 😌

  • Instead of remembering two separate boolean flags, users now have one clearer precision setting: quantize.
  • This makes export commands easier to read, teach, automate, and maintain.

• More future-proof deployment workflows 🔮

  • The old half/int8 flags were too limited for newer precision schemes.
  • The new design opens the door to richer hardware-specific export options without adding more confusing flags later.

• Safer upgrade path ✅

  • Backward compatibility is preserved, so current projects using half or int8 should keep working while users gradually migrate.

• Better results in multi-GPU setups 🖥️

  • Teams validating models across multiple GPUs should now get more accurate confusion matrix visualizations.

• More stable training for OBB models 🎯

  • Users working with oriented bounding boxes, especially small objects, may see fewer training failures and more reliable optimization.

• Fewer frustrating network hangs ⏱️

  • HUB-related actions and downloads should fail more gracefully instead of getting stuck indefinitely.

• Improved docs and onboarding 📚

  • Several updates make deployment, tracking, segmentation guides, and Platform uploads easier to understand for both new and experienced users.

Overall, v8.4.80 is a quality-of-life and deployment-focused release: not a flashy new model drop, but a meaningful improvement for anyone exporting YOLO models to production 🚀

What's Changed

• Clarify tracking persistence and tracker selection by @joeydufourd in https://github.com/ultralytics/ultralytics/pull/24889
• Restore MLflow integration tests on modern mlflow by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/24728
• Gather confusion matrix across GPUs in DDP validation by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/24803
• Forward socket timeout in hub smart_request by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/24743
• Fix double assignment and destructive pop in isolating segmentation guide by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/24701
• Document browser video codec support for Platform dataset uploads by @laodouya in https://github.com/ultralytics/ultralytics/pull/24802
• Add timeout to requests calls in Hub and downloads by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/24734
• Fix NaN OBB loss due to sub-stride boxes by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/24933
• Replace int8 and half args with new quantize by @onuralpszr in https://github.com/ultralytics/ultralytics/pull/24918

New Contributors

• @joeydufourd made their first contribution in https://github.com/ultralytics/ultralytics/pull/24889

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.79...v8.4.80

🌟 Summary

v8.4.79 is a performance-focused release led by a major upgrade to semantic segmentation exports 🚀, plus several quality-of-life improvements in deployment, docs, and developer tooling.

📊 Key Changes

• Big semantic segmentation export boost ⚡
  The standout update from PR #24598 by @onuralpszr makes ONNX and TFLite semantic segmentation exports much faster and lighter by baking the class selection step directly into the exported model.

  • Exported models now output compact class maps instead of large raw score tensors.
  • Reported output size can drop from about 159 MB to 2 MB 📉
  • Inference can be up to 10× faster for this use case.

• Better semantic export compatibility 🧠
  Runtime and postprocessing were updated so exported semantic models work correctly whether they return:

  • traditional floating-point logits, or
  • the new precomputed integer class maps.

• Smaller and simpler similarity search setup 🔍
  PR #24931 by @glenn-jocher removes heavier optional dependencies from the solutions extra:

  • FAISS was replaced with a NumPy-based search approach
  • openai-clip is no longer bundled there as an install dependency
    This makes installs leaner while keeping semantic image search working.

• Tracking code cleanup with no output changes 🛠️
  PR #24930 by @glenn-jocher simplifies tracker internals by removing duplicated Kalman update logic and reusing helper functions.

  • Important note: ByteTrack, BoT-SORT, and OC-SORT behavior stays effectively unchanged.

• Clearer validation for RegionCounter 🛡️
  PR #24890 adds a more understandable error when a region polygon has too few points, making debugging easier for users building region-based counting apps.

• Large documentation refresh across deployment and workflow guides 📚
  Many merged PRs improved guides for:

  • AzureML, Modal, Triton, Raspberry Pi, DeepStream, DALI, and ROS
  • preprocessing, evaluation, fine-tuning, deployment choices, monitoring, and troubleshooting
    These changes don’t introduce major runtime features, but they make setup and usage more accurate and easier to follow.

🎯 Purpose & Impact

• Faster semantic segmentation on edge and mobile 📱
  The main export change is especially valuable for ONNX and TFLite deployments, where output size and postprocessing overhead matter a lot.

• Lower memory and bandwidth costs 💾
  By returning compact class IDs instead of huge dense outputs, semantic models become much more practical for embedded devices, mobile apps, and production pipelines.

• Simpler deployment pipelines 🚚
  Baking argmax into the exported model means less custom postprocessing code outside the model, which reduces integration complexity and potential errors.

• Lighter installs for solution users 🪶
  The dependency cleanup should make optional solution environments easier to install and maintain, especially in constrained or CI environments.

• Improved reliability and usability ✅
  Better validation messages, cleaned-up internals, and more accurate docs help both new and advanced users work faster with fewer confusing issues.

Overall, this release is most important for anyone using semantic segmentation with ONNX or TFLite—those users should see the biggest real-world gains in speed, output size, and deployment simplicity 🎉

What's Changed

• Fix model reference artifacts and restructure defining project goals guide by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/24797
• Refactor Modal, AzureML, and Triton deployment guides by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/24928
• Refactor computer vision project lifecycle guides by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/24921
• Update Raspberry Pi, DeepStream, DALI, and ROS deployment guides by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/24925
• Refactor inference and troubleshooting guides by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/24923
• Restructure monitoring solutions guides and remove boilerplate sections by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/24911
• Fix metrics API usage and warmup_epochs docs in model evaluation guide by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/24904
• Correct normalization and augmentation docs in data preprocessing guide by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/24902
• Complete Google-style docstrings for NumPy helper functions by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24929
• Deduplicate tracker Kalman update and box helpers by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24930
• Standardize Mermaid diagram colors across docs by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24932
• Remove faiss-cpu, openai-clip, and setuptools pins from solutions extra by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24931
• Add https://youtu.be/eZvZEzslTz8 to docs by @RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/24938
• Add Codecov badges to CI docs by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24941
• Clarify Platform dataset ingest API flow by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24943
• Raise clear error for region counter regions with fewer than 3 points by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/24890
• Optimize semantic class-map exports by @onuralpszr in https://github.com/ultralytics/ultralytics/pull/24598

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.78...v8.4.79

🌟 Summary

Ultralytics v8.4.78 is mainly a lightweighting and usability release 🚀—the biggest change removes SciPy from core dependencies, making installs smaller and simpler with no expected behavior changes, while docs were also improved for region counting, predictor argument persistence, and site navigation.

📊 Key Changes

• Removed SciPy from core dependencies 🪶

  • PR #24927 by @glenn-jocher replaces the remaining SciPy-based operations with NumPy-based equivalents.
  • This includes an internal replacement for linear_sum_assignment, used in important matching logic across:
    • RT-DETR
    • SAM / mask matching
    • validation
    • tracking utilities
  • The new implementation was validated against SciPy to machine precision, so this is intended as a drop-in change without functional differences.

• Smaller and cleaner installation footprint 📦

  • Since SciPy was one of the heaviest required packages, removing it reduces the base install size noticeably.
  • This can make setup faster and easier, especially in constrained or cloud environments.

• Region counting documentation was reworked 🎯

  • PR #24816 by @raimbekovm refreshes the RegionCounter guide around actual usage.
  • It now better explains:
    • how region counting works frame by frame
    • how to define one or multiple regions
    • how to access counts in code
    • the difference between live occupancy and cumulative counting

• Clarified model.embed() / predictor argument persistence 📝

  • PR #24926 by @raimbekovm adds an important note that a loaded model may reuse the same predictor across calls.
  • This means arguments like embed can persist unexpectedly unless reset, such as with embed=None.

• Large documentation link cleanup across the docs site 🔗

  • PR #24915 by @raimbekovm converts many internal docs links from absolute URLs to relative links.
  • This helps docs work better in:
    • local builds
    • GitHub previews
    • translated/localized documentation

🎯 Purpose & Impact

• Easier installs for everyone ⚡

  • Removing SciPy lowers dependency weight, which can help users install Ultralytics faster and with fewer environment issues.

• Better portability and deployment 🌍

  • Smaller dependencies are especially useful for CI, Docker images, edge devices, notebooks, and cloud workflows.

• No intended model behavior changes ✅

  • The SciPy removal is an infrastructure improvement, not a model redesign, so users should get the same results with a leaner package.

• Improved reliability for advanced workflows 🤖

  • Because the replaced matching logic is used in areas like RT-DETR, SAM, tracking, and validation, this update keeps those features working while reducing external dependency burden.

• Fewer documentation gotchas 📚

  • The updated RegionCounter guide should make multi-zone counting easier to understand and adopt.
  • The predictor persistence note may save users from confusing behavior when mixing model.embed() and model.predict().

• Better docs experience across languages and previews 🌐

  • Relative linking improves navigation and helps non-English readers stay within the correct localized docs flow.

Overall, v8.4.78 is a quality-of-life release 💡: lighter installs, cleaner internals, and clearer documentation, with the SciPy removal being the standout improvement.

What's Changed

• Use relative cross-file links across docs pages by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/24915
• Document embed argument persistence on the reused predictor by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/24926
• Restructure region counting guide around RegionCounter usage by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/24816
• Remove scipy from core dependencies by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24927

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.77...v8.4.78

🌟 Summary

Ultralytics v8.4.77 is headlined by new knowledge distillation support 🎓 for YOLO training, helping smaller models learn from larger ones to boost accuracy without slowing down inference, alongside a useful RegionCounter crash fix and clearer solution docs 📚.

📊 Key Changes

• 🧠 Knowledge distillation added for YOLO training

  • New training support lets a smaller student model learn from a larger teacher model using the new distill_model argument.
  • A new dis setting controls how strongly the teacher guides training.
  • Added a new internal DistillationModel wrapper to handle:
    • frozen teacher models
    • trainable student models
    • feature matching between models
    • combined standard + distillation loss
  • Includes docs, tests, resume-training support, and checkpoint handling.

• 🚀 New Knowledge Distillation guide

  • Added a dedicated guide with:
    • Python and CLI quickstarts
    • recommended YOLO26 teacher/student pairs
    • usage notes for detect, segment, pose, and OBB
    • benchmark tables for distilled YOLO26 models
    • FAQs and training tips

• 📈 Potential accuracy gains for compact YOLO26 models

  • Release notes show improved validation results for distilled YOLO26 models, while keeping the deployed model size and runtime essentially the same as the original student model.

• 🛠️ RegionCounter crash fixed

  • Fixed a bug where regions added with add_region() could crash during processing because required polygon preparation data was missing.
  • Regions are now fully prepared when added, making RegionCounter more reliable.

• 📚 Documentation refresh for real-time solutions

  • Several YOLO26 guides were reorganized to be easier to follow, including:
    • analytics
    • object counting
    • object blurring
    • object cropping
    • speed estimation
    • distance calculation
  • The analytics guide now also documents area plots.

• 🔎 Documentation SEO/title improvements

  • Fixed duplicate page title issues for some dataset and Roboflow-related docs pages, helping users find the right content more easily.

🎯 Purpose & Impact

• 🎯 Better small-model performance

  • The biggest change is knowledge distillation: it helps smaller YOLO models become more accurate by learning from stronger models during training.
  • This is especially useful for users who want fast, lightweight deployment on edge devices or production systems.

• ⚡ No extra inference cost

  • Distillation improves training, not deployment overhead. That means users can often get a better student model without making inference slower.

• 🧪 More robust training workflows

  • Support for tests, resume behavior, and cleaner checkpoint handling makes distillation more practical for real-world experiments and longer training jobs.

• 🧰 More stable solution usage

  • The RegionCounter fix removes a frustrating crash path, especially for users creating regions programmatically instead of only through predefined setup.

• 📖 Easier onboarding

  • The updated guides make common computer vision tasks more approachable for new users, while also giving experienced users clearer references and cleaner examples.

• 🌍 Better discoverability

  • The docs metadata/title fixes improve site organization and search visibility, which helps users reach the right Ultralytics resources faster.

If you'd like, I can also turn this into a short release note version or a more technical developer-focused summary.

What's Changed

• fix(docs): resolve duplicate title tags on Roboflow and dataset overview pages by @miles-deans-ultralytics in https://github.com/ultralytics/ultralytics/pull/24917
• Restructure real-time solution guides and condense duplicated FAQ examples by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/24910
• Fix RegionCounter crash when regions added via add_region by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/24892
• Add knowledge distillation for YOLO detection models by @lmycross in https://github.com/ultralytics/ultralytics/pull/24894

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.76...v8.4.77

🌟 Summary

Ultralytics v8.4.76 is headlined by a powerful new multi-dataset fine-tuning workflow 🚀, making it much easier to train one YOLO model across several datasets in one run, while also improving tracking defaults, segmentation/tracking stability, and docs clarity.

📊 Key Changes

• 🆕 Multi-dataset fine-tuning via MultiTrainer

  • You can now pass a list or tuple of datasets to model.train().
  • Ultralytics will fine-tune the same base model separately on each dataset in sequence.
  • It returns a per-dataset metrics dictionary and saves a cross-dataset results chart.
  • This expands the old RF100-specific workflow into a more general tool that works for RF100, custom dataset collections, and broader benchmarking.

• 📈 Better multi-dataset benchmarking outputs

  • Training runs are grouped into a shared multitrain directory.
  • Results now include:
    • per-dataset metrics
    • a summary JSON file
    • a bar-chart image comparing results across datasets

• 🔗 Easier Ultralytics Platform dataset/model loading

  • Platform web URLs can now be pasted directly and are automatically converted to ul:// format behind the scenes.
  • This helps streamline workflows using the Ultralytics Platform 🌐

• 🎯 Default tracker changed to tracktrack.yaml

  • TrackTrack is now the default multi-object tracker instead of BoT-SORT.
  • Users who rely on default settings may notice different tracking behavior or ID consistency.

• 🛠️ Segmentation + ReID compatibility fix

  • A fix prevents valid segmentation detections from being incorrectly filtered out when native ReID is enabled.
  • This improves reliability for advanced tracking workflows.

• 📦 DDP training metrics recovery improved

  • Training now better returns metrics in distributed training setups, rather than losing them in some multi-GPU cases.

• 🔄 Revert of recent predict() path-list loading change

  • A previous change meant to reduce memory issues for large file-path lists was reverted because it introduced breaking behavior for exported models with static batch sizes.
  • This restores previous compatibility, especially for ONNX, TensorRT, MNN, and TorchScript export use cases.

• 🧩 Solutions improvements

  • Added show_boxes to instance segmentation solutions so users can turn bounding boxes on or off.
  • Fixed RegionCounter so counts no longer remain stale on empty frames.
  • Clarified region argument docs to better match actual behavior.

• 📚 Documentation updates

  • Added ADE20K semantic segmentation benchmarks for YOLO26 models.
  • Added lightweight COCO16 / COCO32 / COCO64 test subsets to docs.
  • Tracking docs now highlight newer trackers like OC-SORT, Deep OC-SORT, FastTracker, and TrackTrack.
  • Large docs refresh for SEO, consistency, titles, and clearer guides.

🎯 Purpose & Impact

• For researchers and advanced users 🧪

  • The new multi-dataset training feature is the biggest change in this release.
  • It makes it much easier to compare how a single YOLO base model performs across many datasets without building custom scripts.
  • This is especially useful for benchmarking, transfer learning, dataset sweeps, and robustness testing.

• For teams managing many datasets 📂

  • You can now automate repeated fine-tuning runs in a cleaner, more reproducible way.
  • The saved summary plots and JSON outputs make results easier to review, share, and integrate into reporting pipelines.

• For tracking users 🚗

  • The switch to TrackTrack as the default tracker may improve out-of-the-box tracking quality, but it could also slightly change results in existing workflows that depended on default behavior.
  • If you want the old default, you can still set tracker=botsort.yaml explicitly.

• For deployment users ⚙️

  • The revert to the earlier predict() behavior helps avoid export/runtime issues, particularly when using static-batch exported models.
  • This favors compatibility and stability over the recent loader behavior change.

• For solution users 👀

  • Visualization controls are better, and region counting is more accurate frame-to-frame.
  • These fixes reduce confusion and improve reliability in real-world video analytics setups.

• For the broader community 📘

  • Many docs were cleaned up and expanded, making Ultralytics features easier to discover and use.
  • The added ADE20K benchmarks and smaller COCO subsets are especially helpful for quick testing and semantic segmentation evaluation.

In short: v8.4.76 is mainly about making multi-dataset fine-tuning much more practical and scalable 🌟, while also improving default tracking behavior, restoring inference compatibility, and polishing the overall user experience.

What's Changed

• bump ultralytics-inference version to 0.0.22 in documentation by @onuralpszr in https://github.com/ultralytics/ultralytics/pull/24886
• Add show_boxes argument to instance segmentation solution by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/24835
• Fix region argument type and default in solutions docs tables by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/24818
• docs: unified SEO-consistent page titles (≤60, "| Ultralytics" suffix) by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24888
• Fix stale region counts on empty frames in RegionCounter by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/24819
• Revert "Fix OOM when passing large list of file paths to predict()" by @onuralpszr in https://github.com/ultralytics/ultralytics/pull/24895
• Use tracktrack as default tracker by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/24897
• Bump actions/cache from 5 to 6 in /.github/workflows by @dependabot[bot] in https://github.com/ultralytics/ultralytics/pull/24909
• docs: add ADE20K semantic segmentation benchmarks by @lmycross in https://github.com/ultralytics/ultralytics/pull/24901
• Add COCO16/COCO32/COCO64 subset datasets to detect docs table by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24912
• docs: add new tracking tips section for OC-SORT, Deep OC-SORT, FastTracker, and TrackTrack by @onuralpszr in https://github.com/ultralytics/ultralytics/pull/24908
• Restructure five documentation guides and update descriptions by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/24913
• Skip empty mask filter if native ReID is enabled by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/24907
• Correct ParkingManagement arguments in parking management guide FAQ by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/24905
• Add multi-dataset fine-tuning via MultiTrainer by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24914

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.75...v8.4.76

🌟 Summary

Ultralytics v8.4.75 delivers an important CoreML reliability and speed fix for macOS 🍎⚡: CoreML models now run on Apple’s Neural Engine by default instead of using a setting that could crash Python processes on Mac hosts.

📊 Key Changes

• 🚑 Major CoreML backend fix for macOS

  • The CoreML backend now loads models with ComputeUnit.CPU_AND_NE instead of the previous default behavior.
  • This avoids a known macOS coremltools issue where ComputeUnit.ALL or GPU-enabled paths could trigger a hard crash with Error: MLIR pass manager failed.

• ⚡ Neural Engine enabled by default on supported Macs

  • On macOS 13 and newer, CoreML inference now uses the CPU + Neural Engine path automatically.
  • This gives much better performance than CPU-only execution.

• 🛡️ Compatibility fallback for older macOS versions

  • If CPU_AND_NE is not supported, Ultralytics now falls back to CPU_ONLY rather than failing.

• 📝 Documentation updated

  • The CoreML integration docs now explain the new macOS behavior and why avoiding the GPU path currently matters for stability.

🎯 Purpose & Impact

• 🍏 Fixes a serious usability issue for Mac users

  • Before this release, running a CoreML .mlpackage from Python on macOS could crash outright.
  • After this update, CoreML inference should work out of the box on supported Macs.

• 🚀 Improves inference speed

  • The reported result shows about 2.5 ms on the Neural Engine vs 8.5 ms on CPU, roughly a 3× speedup.
  • This is especially valuable for real-time or interactive applications.

• 🔧 Makes CoreML deployment more dependable

  • Users exporting YOLO models to CoreML for local Mac inference should see a much smoother experience with fewer platform-specific failures.

• 👥 Broad impact for Python users on Apple Silicon

  • Anyone using Ultralytics CoreML models from Python on a Mac benefits, especially those working with YOLO26 and other exported .mlpackage models.

• ✅ No major new model architecture changes

  • This release is mainly a backend stability and performance update, not a new model release.
  • The biggest win is that existing CoreML workflows on macOS should now be both faster and far more reliable.

In short: v8.4.75 is a small but high-impact release 🎉—especially for macOS users running CoreML models locally, where it turns a crash-prone path into a fast, working default.

What's Changed

• Run CoreML on the Neural Engine (CPU_AND_NE) on macOS hosts by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24885

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.74...v8.4.75

🌟 Summary

Ultralytics v8.4.74 focuses on more reliable model export and quantization 🔧—especially fixing INT8 export stability on affected GPU setups and preventing flaky OpenVINO export failures on NMS-enabled models.

📊 Key Changes

• 🚨 INT8 calibration now always runs on CPU during ModelOpt export

  • In the most important change from PR #24884 by @glenn-jocher, INT8 calibration for ONNX export was changed to run unconditionally on the CPU execution provider.
  • This replaces the earlier GPU/RTX detection logic, which was found to be unreliable in real-world use.
  • The previous approach could still trigger core dumps or uncatchable crashes on some systems due to TensorRT execution provider behavior and cuDNN ABI mismatches.

• ✅ Safer INT8 export behavior across hardware environments

  • The update avoids calibration on CUDA/TensorRT during the calibration step.
  • Since calibration scales are execution-provider independent, the final INT8 engine remains effectively the same.
  • The tradeoff is simple: slightly slower one-time calibration, but much better export stability.

• 🛠️ Fixed intermittent OpenVINO export failures for NMS models

  • PR #24883 by @glenn-jocher fixes a non-deterministic OpenVINO export issue affecting models exported with nms=True.
  • Previously, OpenVINO could fail during conversion with trace-check errors like “Graphs differed across invocations!”
  • The exporter now pre-traces the model before passing it to OpenVINO, avoiding OpenVINO’s internal retracing behavior that caused random failures.

🎯 Purpose & Impact

• More dependable INT8 exports on RTX and mixed-library environments 💪

  • Users exporting INT8 models should see fewer crashes and failed exports, especially on systems where CUDA, TensorRT, and cuDNN versions interact badly.
  • This is particularly valuable for production pipelines and automated export workflows.

• Better stability is prioritized over calibration speed ⚖️

  • Calibration may take a bit longer because it now runs on CPU only.
  • But this slowdown happens only during the export/calibration step, not during normal model inference.
  • For most users, that is a worthwhile tradeoff for a much more reliable export process.

• OpenVINO exports become more consistent 📦

  • Users working with OpenVINO, especially with NMS-enabled models, should experience fewer random export failures.
  • This helps both local users and teams using the <a href="https://platform.ultralytics.com">Ultralytics Platform</a> for export and deployment workflows.

• Overall release theme: reliability and smoother deployment 🚀

  • This release does not introduce a major new model, but it meaningfully improves the path from training to deployment.
  • If you export models to INT8 or OpenVINO, v8.4.74 should feel safer, more predictable, and easier to trust.

What's Changed

• Fix non-deterministic OpenVINO export failure on NMS models by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24883
• Calibrate INT8 on CPU unconditionally (cuDNN-ABI-safe) by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24884

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.73...v8.4.74

🌟 Summary

Ultralytics v8.4.73 is a stability-focused release that mainly improves INT8 TensorRT export reliability on RTX GPUs 🛠️⚡, while also fixing a memory issue in predict(), improving input order handling, and polishing examples and docs.

📊 Key Changes

• Major fix: safer INT8 calibration on RTX GPUs 🚀

  • The most important update is from PR #24878 by @glenn-jocher.
  • When exporting TensorRT engines with int8=True, RTX cards using the NvTensorRTRTXExecutionProvider now perform calibration on the CPU instead of cuda:0.
  • This avoids crashes caused by execution provider conflicts and cuDNN compatibility mismatches during calibration.
  • Important detail: this does not change the final INT8 engine quality or behavior—only the calibration path is changed to make export more reliable.

• Fix for out-of-memory errors when predicting from large file-path lists 💾

  • PR #24866 fixes a case where passing a large Python list of image paths to model.predict() could load too much into memory and trigger GPU OOM.
  • Local path lists are now handled lazily, so batching works as expected instead of trying to process everything at once.

• Prediction input order is now preserved 📂

  • File lists passed in by users are no longer auto-sorted before loading.
  • This means results now follow the same order as the input list, which is especially helpful for pipelines that depend on exact input-output alignment.

• Better error handling in the OpenCV ONNX example 🖼️

  • PR #24872 adds a clear FileNotFoundError if an image path is missing or unreadable.
  • This replaces a more confusing later-stage failure.

• Small documentation and comment cleanups ✍️

  • PR #24881 improves wording in YOLO-World docs and SAM comments.
  • PR #24821 fixes the QueueManager doc example so copied code uses the correct region format.
  • These changes do not alter runtime behavior, but they make docs easier to follow.

• CI installation step made more resilient 🔁

  • The release also adds retries to a dependency installation step in CI, helping reduce flaky build failures.

🎯 Purpose & Impact

• More dependable INT8 export on RTX hardware ✅

  • Users exporting TensorRT INT8 models on some RTX GPUs should see fewer crashes and failed exports.
  • This is especially valuable for deployment workflows where export stability matters more than raw calibration speed.

• No model accuracy change from the RTX calibration fix 🎯

  • The calibration now runs on CPU for affected RTX setups, but the resulting INT8 engine should remain effectively the same.
  • In short: same intended output, fewer headaches.

• Better scalability for large inference jobs 📈

  • If you run predict() on long image lists, memory use should be much more manageable.
  • This can prevent unnecessary GPU OOM issues and make large batch workflows more practical.

• More predictable automation workflows 🤖

  • Preserving user-supplied file order helps when mapping predictions back to source files in scripts, apps, and data pipelines.

• Friendlier debugging and onboarding 🙌

  • Clearer example errors and cleaner docs reduce friction for newer users and save time for experienced developers.

Overall, v8.4.73 is less about new models and more about making export, inference, and examples more robust and predictable—with the standout improvement being INT8 TensorRT export stability on RTX cards ⚡

What's Changed

• fix minor typos and clarify comments by @fbertil in https://github.com/ultralytics/ultralytics/pull/24881
• Guard against cv2.imread returning None in OpenCV ONNX example by @orbisai0security in https://github.com/ultralytics/ultralytics/pull/24872
• Fix QueueManager docstring region example by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/24821
• Fix OOM when passing large list of file paths to predict() by @leeburt in https://github.com/ultralytics/ultralytics/pull/24866
• Calibrate INT8 on CPU for RTX (NvTensorRTRTX) cards by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24878

New Contributors

• @fbertil made their first contribution in https://github.com/ultralytics/ultralytics/pull/24881
• @leeburt made their first contribution in https://github.com/ultralytics/ultralytics/pull/24866

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.72...v8.4.73

🌟 Summary

v8.4.72 is a small but important stability release that mainly fixes a TensorRT INT8 export crash on some RTX GPUs 🚀, while also improving export environment reliability and cleaning up docs/CI.

📊 Key Changes

• Fixed TensorRT INT8 export crashes on certain RTX cards 🔧

  • The most important change in this release is from PR #24876 by @glenn-jocher.
  • Exporting models with format="engine", int8=True could fail on some RTX GPUs, including cases where ONNX Runtime exposed both TensorRT execution providers.
  • Ultralytics changed the INT8 calibration step to use CUDA with CPU fallback instead of triggering conflicting TensorRT provider combinations.

• Improved TensorRT/ONNX export image compatibility for CUDA 12 🐳

  • PR #24877 pins onnxruntime-gpu to below 1.27.0 in the export Docker image.
  • This avoids breakage caused by newer ONNX Runtime GPU builds that now expect CUDA 13, while the export image still uses CUDA 12.8.

• GitHub Actions checkout updated from v6 to v7 ⚙️

  • PR #24873 refreshes CI workflows to a newer actions/checkout version.
  • This is a maintenance update for the project’s automation pipeline.

• Documentation updates and cleanup 📚

  • PR #24863 refreshes the Queue Management guide with a newer YOLO26 video and updated wording.
  • PR #24875 simplifies the Triton C++ example README by removing extra contributor footer content.
  • PR #24851 adds clearer page titles for account settings and YOLO configuration docs, helping site clarity and SEO.

🎯 Purpose & Impact

• More reliable INT8 TensorRT exports on RTX hardware 💡

  • Users exporting optimized TensorRT engines should see fewer failures on affected NVIDIA RTX systems.
  • This is especially helpful for developers deploying compact, fast INT8 inference pipelines.

• Better support for modern RTX environments 🖥️

  • The fix avoids a low-level ONNX Runtime provider conflict that was causing exports to crash before the model was even fully built.
  • In simple terms: INT8 export now works more reliably on hardware that previously broke unexpectedly.

• Safer Docker-based export workflows 🛡️

  • Pinning onnxruntime-gpu prevents version mismatches that could break ONNX inference and export tests inside CUDA 12 environments.
  • This should make CI and container-based deployment setups more predictable.

• No major model architecture changes 📌

  • This release does not introduce a new model or training feature.
  • The focus is on export stability, environment compatibility, and documentation polish.

• Practical benefit for users ✅

  • If you use Ultralytics for training and then export to TensorRT INT8 for production, this release is worth adopting.
  • If you mainly use standard training or FP16 export, the impact is smaller but still positive thanks to general reliability improvements.

What's Changed

• Add https://youtu.be/TEVPiGCxB0o to docs by @RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/24863
• docs: small cleanup triton contributions by @onuralpszr in https://github.com/ultralytics/ultralytics/pull/24875
• Pin onnxruntime-gpu<1.27.0 for CUDA 12 export image by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24877
• Bump actions/checkout from v6 to v7 in /.github/workflows by @UltralyticsAssistant in https://github.com/ultralytics/ultralytics/pull/24873
• Fix duplicate title tags on Spanish account settings and YOLO config pages by @miles-deans-ultralytics in https://github.com/ultralytics/ultralytics/pull/24851
• Fix TensorRT INT8 export crash on RTX cards exposing the NvTensorRTRTX EP by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24876

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.71...v8.4.72

🌟 Summary

Ultralytics v8.4.71 is mainly a release-tag update 📦, but the most important underlying change is a major refactor of the official C++ examples that makes YOLO deployment in C++ much more unified, easier to use, and more capable across backends 🚀

📊 Key Changes

• Release bump to 8.4.71 🔖

  • The current PR by @glenn-jocher primarily updates the package version from 8.4.70 to 8.4.71.
  • No direct Python API, model architecture, or runtime behavior changes are introduced in this version-bump PR itself.

• Big C++ examples refactor landed in this release 🛠️

  • C++ inference examples were reorganized into a single examples/cpp/ structure instead of being scattered across many separate folders.
  • Multiple backends are now aligned under one shared design:
    • OpenCV-DNN
    • ONNX Runtime
    • LibTorch
    • MNN
    • OpenVINO
    • Triton
  • Shared header-only utilities were added for:
    • task detection
    • post-processing
    • rendering/annotation
    • CLI handling
    • COCO fallback names

• Much broader C++ task support 🎯

  • The refactored examples now support all major YOLO tasks across C++:
    • detection
    • segmentation
    • pose
    • OBB
    • classification
    • YOLO26 semantic segmentation
  • They also support multiple model generations, including YOLOv8, YOLO11, and YOLO26.

• Automatic task and model behavior detection 🤖

  • Many C++ examples can now automatically determine the task from model metadata or output shapes.
  • This reduces manual setup and makes examples easier to reuse with different exported models.

• More consistent C++ developer experience 🧰

  • Unified build targets like yolo_<backend>
  • Standardized CLI arguments such as --model and --source
  • Cleaner READMEs and better docs across examples
  • Default examples now commonly point to lightweight YOLO26 models

• Ultralytics Platform GPU docs expanded ☁️

  • Documentation now includes two new Blackwell GPU options on the Ultralytics Platform:
    • RTX PRO 4000 Blackwell with 24 GB VRAM
    • RTX PRO 5000 Blackwell with 48 GB VRAM
  • Platform cloud GPU counts were updated accordingly across docs.

🎯 Purpose & Impact

• For most users: this release is mostly about packaging and delivery** ✅

  • v8.4.71 helps tools and environments track the newest Ultralytics build correctly.
  • It improves release consistency without forcing disruptive API changes.

• For C++ developers: this is the real headline** 💡

  • The refactor makes C++ deployment much easier to understand and maintain.
  • Instead of learning different example layouts for each backend, users get a more unified workflow.
  • This should lower friction for deploying YOLO models in production C++ systems.

• For teams using multiple model types: better flexibility** 🔄

  • Supporting more tasks and generations in a common C++ structure means fewer custom modifications are needed when switching models.
  • That can save engineering time and reduce mistakes.

• For YOLO26 adoption: stronger deployment support** 🚀

  • Adding clearer C++ coverage for YOLO26, including semantic segmentation, helps bring the latest recommended Ultralytics model into more real-world applications.

• For Ultralytics Platform users: more cloud hardware choices** ⚡

  • The added Blackwell GPU options expand available training hardware in the docs and platform guidance.
  • This can help users choose a better balance of performance, VRAM, and cost.

In short: the tag PR itself is just a version bump, but v8.4.71 packages a meaningful C++ usability upgrade and refreshed platform GPU documentation 🎉

What's Changed

• Add RTX PRO 5000 and RTX PRO 4000 Blackwell to Platform GPU docs by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24868
• Refactor and unify C++ inference examples 🐪 by @onuralpszr in https://github.com/ultralytics/ultralytics/pull/24867
• Official C++ Examples refactor 🐪 by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24869

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.70...v8.4.71

🌟 Summary

v8.4.70 is a reliability-focused release 🛠️ led by an important fix for multi-GPU training, plus safer checkpoint loading, a plotting fix for segmentation results, and clearer docs around YOLO26 heatmaps, training outputs, and deployment 📚

📊 Key Changes

• 🐛 Fixed DDP AMP synchronization in multi-GPU training in PR #24810 by @LZ-QWQ

  • This is the main update in the release.
  • It fixes a bug where AMP mixed-precision training could be disabled on one GPU rank but remain enabled on others during Distributed Data Parallel training.
  • After the fix, all ranks now correctly share the same AMP setting.

• ✅ Safer checkpoint loading with clearer errors in PR #24857 by @glenn-jocher

  • Ultralytics now gives a clear, user-friendly error when loading unsupported checkpoint formats, such as plain state_dict files or some older YOLOv5-style module references.
  • This replaces confusing crashes like AttributeError or KeyError.

• 🎨 Fixed Results.plot() for masks-only outputs in PR #24635 by @ahmet-f-gumustas

  • Segmentation results that contain masks but no bounding boxes no longer crash during plotting.
  • This helps users working with custom or partial result objects.

• 🔥 Heatmaps guide was rewritten around real tracking use cases in PR #24815 by @raimbekovm

  • The documentation now better explains that YOLO26 heatmaps are for object tracking over video, not just generic data visualization.
  • It clarifies real-world uses like traffic flow, crowd movement, and region-based counting.

• 📁 Documented the save_dir training argument in PR #24831 by @raimbekovm

  • Users can now more easily understand how to send training outputs to an exact folder without auto-increment behavior.

• 🧩 Small quality-of-life fixes in docs and config

  • Removed an unused region argument from the instance segmentation guide in PR #24833 by @raimbekovm
  • Corrected figsize typing to floats in Solutions config in PR #24832 by @raimbekovm
  • Fixed doc rendering issues and broken links, including TensorRT quantization docs, in PRs #24822, #24856, and #24858 📘

🎯 Purpose & Impact

• 🚀 More reliable distributed training

  • The AMP/DDP fix is especially important for users training on multiple GPUs.
  • It prevents inconsistent precision behavior across devices, which could otherwise lead to unstable training, debugging headaches, or hard-to-trace performance issues.

• 🛡️ Fewer cryptic failures

  • The checkpoint-loading improvements make model-loading problems much easier to understand.
  • This is especially helpful for users migrating from older formats or using custom-saved weights.

• 🎯 Better support for segmentation workflows

  • The plotting fix makes Ultralytics more robust for segmentation-only results and edge cases.

• 📚 Clearer guidance for real-world YOLO26 usage

  • The heatmaps documentation update makes it easier for a broad audience to understand how YOLO26 can be used for video analytics, such as retail, transportation, and crowd analysis.

• ⚙️ Smoother day-to-day experience

  • While many of the remaining changes are documentation and maintenance updates, they reduce confusion, improve discoverability, and make the platform easier to use overall.

In short: v8.4.70 is mainly a stability and usability release ✅ with the standout improvement being a critical fix for consistent AMP behavior in distributed multi-GPU training.

What's Changed

• Fix TensorRT quantization docs link by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24856
• Remove region from instance segmentation guide arguments table by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/24833
• Fix docs content rendering as raw code blocks by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/24822
• Update formatting, headers, and redirects by @pderrenger in https://github.com/ultralytics/ultralytics/pull/24826
• Refactor heatmaps guide with tracking-focused framing by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/24815
• Document save_dir training argument by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/24831
• Fix figsize type annotation in SolutionConfig by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/24832
• Raise clear error for non-Module checkpoints and YOLOv5 models.yolo by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24857
• Fix AttributeError in Results.plot() for masks-only results by @ahmet-f-gumustas in https://github.com/ultralytics/ultralytics/pull/24635
• Fix docs workflow PR branch sync by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24858
• Fix DDP AMP broadcast state synchronization by @LZ-QWQ in https://github.com/ultralytics/ultralytics/pull/24810

New Contributors

• @LZ-QWQ made their first contribution in https://github.com/ultralytics/ultralytics/pull/24810

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.69...v8.4.70

🌟 Summary

Ultralytics v8.4.69 is a small but important reliability update 🔒 that improves safe checkpoint loading, especially for some YOLO11 and YOLOv8 segmentation, pose, OBB, and classification models.

📊 Key Changes

• 🔐 Safe loading now supports certain getattr references inside checkpoints

  • Some model checkpoints store internal references using getattr, which were previously blocked when safe loading was enabled.
  • This update allows those checkpoints to load correctly without turning off the safety mechanism.

• 🤖 Fixes safe loading for several model types

  • Affects checkpoints used by:
    • YOLO11 segmentation
    • YOLO11 classification
    • YOLO11 pose
    • YOLO11 OBB
    • Similar affected YOLOv8 checkpoint variants

• 🧩 Added restricted support for enum-based references

  • The safe loader now understands certain enum values used in saved checkpoints, such as interpolation settings.

• 🛡️ Still keeps the load process restricted

  • Instead of broadly allowing any getattr use, the change adds a limited internal handler that only permits approved cases.
  • This helps preserve security while improving compatibility.

• 📦 Version bump

  • Package version updated from 8.4.68 to 8.4.69.

🎯 Purpose & Impact

• ✅ Fewer loading/export errors with safe mode enabled

  • Users working with ULTRALYTICS_SAFE_LOAD=1 should now be able to load and export affected checkpoints more smoothly.

• 🔒 Better security without losing usability

  • The goal is to keep safe loading strict, while avoiding unnecessary failures for legitimate Ultralytics model files.

• 🚀 Improves deployment and export workflows

  • This is especially helpful for users exporting models in secure or controlled environments, where safe loading is preferred or required.

• 👥 Good quality-of-life fix for everyday users

  • If you use pretrained segmentation, pose, OBB, or classification checkpoints, this update can reduce friction without changing your training or inference workflow.

• 📌 Low-risk, high-value release

  • There are no major new model architectures or training features here; this release is mainly a stability and compatibility fix with practical impact.

If helpful, I can also provide a one-paragraph changelog version or a more technical developer-focused summary.

What's Changed

• Allow safe load of checkpoints with getattr by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/24852

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.68...v8.4.69

🌟 Summary

Ultralytics v8.4.68 is mainly a training-quality and reliability release, led by a major improvement for semantic segmentation: you can now use class weighting to better handle imbalanced datasets 🎯🧠

📊 Key Changes

• Semantic segmentation now supports class weights via cls_pw 🖼️⚖️

  • Added in PR #24843 by @Y-T-G.
  • This is the headline change in the release.
  • Previously, semantic segmentation did not support user-configurable class weighting.
  • Now, training can emphasize underrepresented classes, which is especially useful when some classes appear far less often than others.

• Default semantic training behavior stays the same 🔒

  • If cls_pw is not set, or is 0, Ultralytics keeps the previous behavior.
  • Existing Cityscapes-specific weighting is still used automatically when appropriate.
  • This means current training pipelines should continue working without surprises.

• Better weighting logic for segmentation 🧮➡️🧠

  • Semantic segmentation now uses a weighting approach based on class pixel frequency that is more suitable for dense pixel-level tasks than the method used in detection.
  • This helps avoid overly extreme weighting while still giving rare classes more importance.

• More robust handling of missing or unreadable labels/masks 🛠️

  • The new class-counting logic safely handles cases where labels or masks are missing or unreadable.
  • Instead of breaking training, it can fall back more gracefully.

• Detection classification loss can now apply class weights too 📦

  • The release also refactors class counting and weighting into reusable trainer helpers.
  • Detection training now applies computed class weights directly in the classification loss.

• Several production stability fixes 🚑

  • PR #24840 by @glenn-jocher fixes four real-world issues:
    • crash from empty prediction batches
    • corrupt cached AMP asset loading
    • truncated downloads
    • invalid non-mapping YAML files
  • These changes improve day-to-day reliability for training and inference.

• Safer model loading on older PyTorch versions 🔐

  • PR #24830 by @glenn-jocher makes restricted loading fall back cleanly on torch<2.5 instead of erroring out.
  • Helpful for older toolchains and export workflows.

• Security fix for similarity search UI 🛡️

  • PR #24828 escapes user query input properly to prevent a high-severity browser injection issue.

• Docs and platform updates 📚

  • Added Windows on Snapdragon QNN benchmarks for YOLO26n across all six tasks, showing strong NPU speedups 🚀
  • Updated Ultralytics Platform GPU pricing docs 💳
  • Large cleanup of broken links, page titles, URLs, and navigation for smoother docs browsing.

🎯 Purpose & Impact

• Better semantic segmentation on imbalanced datasets 🌍

  • If your dataset has rare classes like small road signs, defects, or medical regions, the model can now pay more attention to them during training.
  • This may improve results where minority classes were previously overlooked.

• Safer upgrade path ✅

  • Because default behavior is preserved unless you use cls_pw, most users can upgrade without changing existing results.

• More control for advanced users 🎛️

  • Researchers and practitioners training custom semantic datasets now get a useful tuning option that was missing before.

• Improved robustness in production 🏭

  • Fewer crashes from bad files, corrupt caches, malformed YAML, or odd input batches means smoother automation and more reliable pipelines.

• Broader hardware confidence 💻

  • The new Windows on Snapdragon benchmark documentation helps users evaluate YOLO26 deployment on modern ARM laptops and NPU hardware.

• Better security and compatibility 🔒

  • Safer web templates and graceful fallback behavior on older PyTorch versions reduce friction for both developers and deployers.

Overall, v8.4.68 is a practical quality release with its biggest win being more capable and customizable semantic segmentation training, plus several important fixes that make Ultralytics more stable and dependable in real-world use 🚀

What's Changed

• Skip restricted model loading on torch<2.5 by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24830
• Fix docs broken link checks by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24836
• docs: add Windows on Snapdragon benchmarks by @ShuaiLYU in https://github.com/ultralytics/ultralytics/pull/24834
• Update Platform GPU pricing by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24838
• Fix ClearML project paths by @MUKILAN0608 in https://github.com/ultralytics/ultralytics/pull/24827
• Escape similarity search query by @orbisai0security in https://github.com/ultralytics/ultralytics/pull/24828
• Add https://youtu.be/zF2T17ppKIE to docs by @RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/24837
• Fix four production Sentry issues: predict batch, corrupt AMP cache, truncated downloads, non-mapping YAML by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24840
• Fix docs metadata and URL issues by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24839
• Semantic segmentation class weights by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/24843

New Contributors

• @orbisai0security made their first contribution in https://github.com/ultralytics/ultralytics/pull/24828
• @MUKILAN0608 made their first contribution in https://github.com/ultralytics/ultralytics/pull/24827

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.67...v8.4.68

🌟 Summary

v8.4.67 is a small but meaningful release focused on safer model loading 🔒, plus an important TT100K dataset label fix 🛣️ and a batch of documentation/link cleanups 📚.

📊 Key Changes

• 🔒 New opt-in safe model loading via ULTRALYTICS_SAFE_LOAD (@glenn-jocher, PR #24829)

  • Ultralytics added a new environment flag, ULTRALYTICS_SAFE_LOAD=true, that enables a safer way to load model checkpoints.
  • When enabled, checkpoint loading uses PyTorch’s newer weights_only=True behavior and only rebuilds known Ultralytics and PyTorch model classes from an automatically generated allow-list.
  • This replaces the older unused SafeUnpickler approach.
  • Important: this is off by default, so existing workflows keep working as before.

• 🛣️ TT100K dataset class list corrected to the official 221-category set (@glenn-jocher, PR #24718)

  • The TT100K.yaml dataset config had duplicated and incorrect class names.
  • That caused some class IDs to silently lose labels during conversion, even though training still proceeded.
  • The release now aligns TT100K with the official upstream category list, including missing real classes.

• 📘 Rust inference docs updated to ultralytics-inference 0.0.21 (@onuralpszr, PR #24825)

  • Documentation examples now point to the latest Rust inference crate version.

• 🔗 Many documentation links were refreshed and corrected (@glenn-jocher, PR #24824 and #24817)

  • Updated links across integrations, datasets, examples, and guides.
  • Includes a fix for the CoreML NMS glossary link and many canonical URL updates.

🎯 Purpose & Impact

• 🔐 Safer checkpoint loading for security-conscious users

  • The new opt-in safe loading mode helps reduce risk when loading model files by allowing only recognized model components.
  • This is especially useful for teams working in stricter environments or aligning with newer PyTorch defaults.
  • Because it is optional, users with custom or forked architectures are not disrupted.

• ♻️ Better future compatibility with PyTorch

  • PyTorch 2.6+ is moving toward weights_only loading by default.
  • This update helps Ultralytics stay aligned with that direction while preserving backward compatibility today.

• 🚫 Less chance of silent dataset errors

  • The TT100K fix is a big quality improvement for anyone training on that traffic-sign dataset.
  • Previously, some categories could effectively receive no labels without an obvious crash.
  • Now, dataset conversion and training should be much more trustworthy 🎯

• 🧰 Minimal disruption for most users

  • No major model architecture changes were introduced in this tag.
  • Standard loading behavior remains unchanged unless you explicitly enable safe loading.
  • Most other changes are documentation polish, so the release is low-risk and practical ✅

• 📚 Clearer onboarding and fewer broken docs

  • Updated links and version references make tutorials and integration guides easier to follow, especially for newer users.

If you want, I can also provide a one-paragraph release note version or a developer-focused summary of v8.4.67.

What's Changed

• Fix CoreML NMS glossary link by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24817
• Fix TT100K class names to match the official 221-category list by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24718
• chore: update ultralytics-inference version to 0.0.21 in documentation by @onuralpszr in https://github.com/ultralytics/ultralytics/pull/24825
• Update redirected docs links by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24824
• Add opt-in weights_only model loading via ULTRALYTICS_SAFE_LOAD by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24829

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.66...v8.4.67

🌟 Summary

Ultralytics v8.4.66 improves built-in NVIDIA GPU monitoring reliability 🔋, adds better semantic segmentation class filtering and metrics 🧠, and enhances CoreML semantic exports on iPhone with full-resolution outputs 📱.

📊 Key Changes

• Main update: NVIDIA GPU monitoring is now more reliable 🚀

  • nvidia-ml-py was added as a standard dependency by @glenn-jocher.
  • GPU stats collection now uses NVIDIA’s NVML library directly instead of trying to install support at runtime.
  • This affects built-in GPU reporting in utilities like auto device selection and system logging.

• Semantic segmentation got a meaningful feature upgrade 🎯

  • The classes filter now works properly for semantic segmentation during:
    • dataset loading
    • prediction
    • validation
    • metric reporting
  • Non-selected classes are cleanly ignored instead of being mixed into results.
  • Added safeguards for unsupported cases like single_cls=True in semantic segmentation.

• Semantic segmentation metrics are more accurate 📈

  • mIoU now ignores classes that are not actually present in the ground truth.
  • Per-class reporting was improved so summaries better reflect what was really evaluated.

• CoreML semantic exports now keep full-resolution maps 📱✨

  • CoreML exports for semantic segmentation now output full-resolution semantic maps instead of lower-resolution stride-8 maps.
  • This means much sharper segmentation overlays on Apple devices like iPhones.

• Updated iPhone CoreML benchmark docs 📊

  • Performance numbers were refreshed using optimized profile-mode measurements.
  • Documentation now better reflects real YOLO26 CoreML behavior, especially for semantic models.

• TensorRT 11 export flow was simplified ⚙️

  • The separate half argument was removed from ONNX precision conversion.
  • TensorRT 11 export now has a cleaner path for FP16 and INT8 generation.
  • INT8 calibration preprocessing was streamlined for cleaner and more maintainable export behavior.

• Documentation improvements across YOLO26 guides 📝

  • The Model Testing guide was rewritten for clarity and accuracy.
  • The YOLO26 Training Recipe guide was improved to better explain how official checkpoints were trained.
  • Small fixes were also made to JupyterLab and SAM 3 docs.

• CI and packaging cleanup 🛠️

  • CI dependency installs now retry automatically, reducing flaky failures.
  • Since nvidia-ml-py is now a standard dependency, extra manual installs were removed from CI and Docker setup.

🎯 Purpose & Impact

• More dependable GPU stats for users with NVIDIA hardware ✅
  Users should see fewer issues with built-in GPU monitoring, logging, and hardware-aware utilities because the required NVIDIA package is now installed up front instead of handled on the fly.

• Better usability for semantic segmentation workflows 🎨
  If you work with semantic segmentation and only care about certain classes, filtering now behaves much more intuitively during prediction and evaluation.

• More trustworthy segmentation metrics 📏
  Reported semantic segmentation scores are now less misleading, especially on datasets where some classes are absent or filtered out.

• Sharper CoreML semantic results on iPhone and Apple devices 🍎
  Semantic masks should look much cleaner and less blocky, improving visual quality for mobile deployment.

• Cleaner export experience for TensorRT 11 ⚡
  Developers exporting models for NVIDIA deployment get a simpler and more maintainable precision conversion workflow.

• Clearer docs for a broader audience 📚
  New and experienced users alike should find YOLO26 testing, training, and deployment guidance easier to follow.

Overall, v8.4.66 is a practical quality-focused release: it strengthens GPU monitoring, improves semantic segmentation behavior, and brings better mobile export quality without changing the core user workflow.

What's Changed

• Remove unused half argument and optimize INT8 calibration in TensorRT 11 export by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/24795
• Improve model testing guide structure and fix validation metric claims by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/24800
• Export full-resolution semantic maps to CoreML by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24799
• Refresh iPhone CoreML benchmarks with profile-mode measurements by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24804
• Fix evaluate link target in JupyterLab integration docs by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/24808
• Clarify multiple bounding boxes behavior in SAM3 semantic predictor docs by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/24807
• Add retry wrapper to flaky CI installs by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24809
• Adapt classes filter to semantic segmentation task by @lmycross in https://github.com/ultralytics/ultralytics/pull/24806
• Improve YOLO26 training recipe guide structure and accuracy by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/24805
• Add nvidia-ml-py to pyproject.toml by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/23922

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.65...v8.4.66

🌟 Summary

v8.4.65 is mainly a mobile inference performance release 📱⚡, with the biggest upgrade improving QNN exports for Snapdragon NPUs and speeding up semantic segmentation outputs by moving costly work inside the model graph.

📊 Key Changes

• 🚀 Major QNN export performance upgrade in PR #24790 by @glenn-jocher

  • QNN exports now use channel-last (NHWC) input, which better matches Qualcomm Hexagon NPU hardware and camera-native image layouts.
  • This avoids extra data reordering at inference time, reducing overhead on both the app side and the NPU side.
  • Implemented through new export wrappers like QNNModel instead of fragile post-export graph editing.

• 🧠 Semantic segmentation exports are much faster on-device

  • For QNN and CoreML semantic models, export now embeds ArgMax / class-map generation directly in the graph.
  • Instead of returning large floating-point logits that apps must decode afterward, exported models can now return a compact per-pixel class map directly.
  • New ClassMapModel wrapper handles this behavior during export.

• 📱 Better mobile runtime compatibility

  • Semantic predict/validation code was updated so Ultralytics can correctly handle both:
    • traditional logits outputs
    • new exported class-map outputs
  • This makes the faster export behavior work more cleanly across deployment and evaluation workflows.

• 📉 Measured impact highlighted in docs

  • QNN semantic performance became much more stable, replacing previously erratic decoding times with far more predictable behavior.
  • CoreML semantic exports also saw notable end-to-end speedups from in-graph class-map generation.

• 🛠️ RKNN export improvements

  • RKNN export now properly supports FP16 intermediate ONNX export.
  • Added RKNN INT8 export test coverage to improve reliability.
  • RKNN export docs/tables were updated accordingly.

• 🐳 Docker build reliability improvements

  • Dockerfiles now use more cache-conscious install patterns and extra cleanup steps to reduce disk pressure during builds.
  • The Docker workflow was simplified by removing an extra assistant-trigger step.

• 📚 Docs and export guidance improvements

  • Expanded and refreshed docs for CoreML and QNN, including clearer deployment guidance and performance notes.
  • Export docs now better explain YOLO26 end-to-end detection output formats.
  • Tracking docs now clarify support for OBB models as well.
  • Rust inference docs were updated to ultralytics-inference version 0.0.19.

• ✅ CI and workflow fixes

  • Docs redeploys now trigger more reliably when documentation changes land on main.
  • CUDA training tests are now explicitly skipped on Jetson devices, making CI behavior cleaner and easier to maintain.

🎯 Purpose & Impact

• ⚡ Faster mobile inference on Snapdragon devices
  Users deploying YOLO on Qualcomm hardware should see lower overhead and better real-world efficiency, especially in camera-based apps where image buffers are already channel-last.

• 🧩 Less app-side postprocessing work
  By returning semantic class maps directly from exported models, apps no longer need to spend as much CPU time decoding huge segmentation outputs.

• 📈 Better and more stable semantic segmentation performance
  This is especially important for real-time mobile experiences, where unpredictable postprocessing delays can cause lag or jitter.

• 🔒 More robust export pipeline
  Replacing manual ONNX graph surgery with clean wrapper modules makes exports easier to maintain and less error-prone over time.

• 📱🍎 Benefits extend beyond QNN CoreML semantic exports also gain from the same in-graph class-map idea, so Apple-device deployment gets a speed boost too.

• 🛠️ Improved reliability for deployment workflows Better RKNN testing, cleaner Docker builds, and more dependable docs publishing all reduce friction for developers working across edge and production environments.

• 👥 Broad user impact

  • Mobile and edge developers get the biggest benefit from this release.
  • Semantic segmentation users should see the clearest speed and usability gains.
  • General users also benefit from clearer docs, cleaner CI, and more reliable export behavior.

In short: v8.4.65 is a strong deployment-focused release 🎉, with the standout improvement being faster, more hardware-friendly QNN exports and smarter semantic segmentation outputs for mobile AI.

What's Changed

• Improve skipping training on Jetson Test CIs by @lakshanthad in https://github.com/ultralytics/ultralytics/pull/24557
• Trigger docs redeploys on changes by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24784
• Replace Enterprise '50+ members' with custom team size in Platform docs by @mykolaxboiko in https://github.com/ultralytics/ultralytics/pull/24785
• Normalize YAML docs links by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24788
• Fix docs deploy change detection by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24787
• Improve docs, comments, and correctness by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24789
• Enable RKNN FP16 export and add RKNN INT8 test by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/24762
• Remove Docker workflow assistant trigger by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24783
• docs: ultralytics-inference version to 0.0.19 in rust docs by @onuralpszr in https://github.com/ultralytics/ultralytics/pull/24793
• QNN channel-last + semantic ArgMax performance improvements by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24790

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.64...v8.4.65

🌟 Summary

Ultralytics v8.4.64 makes tracking easier and more reliable 🎯—especially with new auto-downloaded YOLO26 ReID ONNX encoders—while also improving QNN export, multi-GPU training stability, logging, and export robustness 🚀

📊 Key Changes

• Tracking got a major usability upgrade with auto-downloaded YOLO26 ReID ONNX encoders 🤖📦
  You can now use built-in tracking ReID models like yolo26n-reid.onnx through yolo26x-reid.onnx without manually downloading or exporting them first.

• Tracking ReID matching is more reliable when appearance features are missing 🔍
  TrackTrack now treats missing embeddings as “unknown” instead of assuming they are bad matches, so it can fall back to motion cues more intelligently.

• QNN export for Qualcomm devices was significantly improved 📱⚡
  QNN models now export as a single self-contained *_qnn.onnx file instead of a folder, with metadata embedded inside. The release also fixes architecture mapping, improves backend loading compatibility, and updates quantization behavior for better Snapdragon deployment.

• YOLO26x distributed training stability was fixed 🧠🖥️
  A DDP deadlock issue affecting some multi-GPU training runs was resolved by restoring handling for unused parameters in conditional branches.

• MLflow failures no longer crash training 📉🛡️
  If MLflow tracking setup or logging fails, training now continues instead of aborting the run.

• TensorFlow export subprocess calls are safer 🔐
  Edge TPU and TensorFlow.js export commands now avoid shell-based path handling issues, reducing problems with unusual file paths and improving security.

• Progress/logging output was cleaned up 🖥️✨
  Fixed premature 100% progress display and console log duplication issues, especially useful in platform or remote logging environments.

• Version checking is more accurate ✅
  parse_version() now consistently returns 3-part version tuples, fixing incorrect version comparisons like 6.0 vs 6.0.0.

• FP16 quantization/export reliability was improved ⚙️
  ONNX mixed-precision conversion now uses the correct input name and sample input shape, helping TensorRT/ModelOpt workflows work more consistently.

• Docs were refreshed across several areas 📘
  Updates include clearer tracking/ReID docs, QNN docs, Conda install guidance, hyperparameter tuning explanations, TrackZone behavior, custom trainer checkpoint loading, OpenVINO benchmark references, and terminal visualization guidance.

🎯 Purpose & Impact

• Easier multi-object tracking setup 🚀
  The headline change removes a common setup headache: users can now enable tracking ReID with ready-made YOLO26 ONNX encoders directly, making advanced tracking more accessible to both beginners and production teams.

• Better tracking quality in difficult scenes 🎥
  The ReID fallback improvement should reduce bad associations when objects are briefly occluded or appearance features are unavailable.

• Simpler Qualcomm deployment 📱
  QNN export is now easier to manage and deploy thanks to the single-file output format and improved compatibility across Snapdragon targets.

• More dependable training at scale 🖥️
  Teams training larger YOLO26 models on multiple GPUs should see fewer hangs and more stable runs.

• Fewer pipeline interruptions 🛠️
  MLflow and export-related fixes help ensure optional integrations do not stop core training or deployment workflows.

• Cleaner user experience ✨
  Console logging and progress bar fixes make training output easier to trust and monitor, especially in shared platforms and dashboards.

• Lower friction for installation and documentation 📚
  Updated guides make setup, tuning, export, and custom workflows easier to understand and reproduce.

In short, v8.4.64 is a tracking-focused quality-of-life release with especially strong benefits for users working with YOLO26 tracking, Qualcomm QNN deployment, and stable training/export pipelines 🚀

What's Changed

• Fix parse_version to always return a 3-tuple by @bujna94 in https://github.com/ultralytics/ultralytics/pull/24680
• Update OpenVINO benchmarks version with openvino==2026.2.0 by @lakshanthad in https://github.com/ultralytics/ultralytics/pull/24691
• Fix MLflow callback crashing training when tracking backend setup fails by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/24705
• docs: clarify tuning iterations and population size by @xiazl1993 in https://github.com/ultralytics/ultralytics/pull/24706
• Use argv lists for TensorFlow export subprocess calls by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/24767
• Add note on reloading models with custom trainer classes by @Harshxth in https://github.com/ultralytics/ultralytics/pull/24739
• fix(trackers): TrackTrack ReID falls back to motion for missing-feature pairs by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/24768
• Fix DDP deadlock for YOLO26x: re-add find_unused_parameters by @MeVeryHandsome in https://github.com/ultralytics/ultralytics/pull/24620
• Fix premature 100% in TQDM and ANSI clear-line leak in ConsoleLogger by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24773
• Exclude boot/firmware partitions from SystemLogger drive list by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24764
• Fix QNN HTP arch mapping, monolithic EP loading, and A16W8 quantization by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24775
• Specify calibration data shape for FP16 quantization by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/24779
• Improve Conda quickstart guide by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/24781
• Improve terminal inference results guide by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/24778
• Convert QNN export to a single ONNX file by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24780
• Refactor TrackZone guide: fix FAQ example and document zone region behavior by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/24772
• Update conda CUDA install pin to pytorch-cuda=12.1 by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/24782
• Tracker ReID ONNX encoders autodownload by @onuralpszr in https://github.com/ultralytics/ultralytics/pull/24774

New Contributors

• @xiazl1993 made their first contribution in https://github.com/ultralytics/ultralytics/pull/24706
• @bujna94 made their first contribution in https://github.com/ultralytics/ultralytics/pull/24680
• @MeVeryHandsome made their first contribution in https://github.com/ultralytics/ultralytics/pull/24620
• @Harshxth made their first contribution in https://github.com/ultralytics/ultralytics/pull/24739

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.63...v8.4.64

🌟 Summary

🚀 Ultralytics v8.4.63 mainly adds TensorRT 11 export support with FP16 and INT8 quantization via NVIDIA ModelOpt, while also expanding built-in multi-object tracking with several new tracker options and improving reliability, performance, and docs.

📊 Key Changes

• 🧠 Major export upgrade: TensorRT 11 support

  • The standout update in this release is PR #24735 by @onuralpszr.
  • Ultralytics now supports exporting models to NVIDIA TensorRT 11, which had previously broken due to API changes in TensorRT 11.
  • For FP16 and INT8 export on TensorRT 11, Ultralytics now uses NVIDIA ModelOpt instead of the older TensorRT builder flags and calibrator interface.
  • This keeps export working across:
    • TensorRT 7–10 with the legacy path
    • TensorRT 11 with the new strongly-typed ModelOpt path
  • Tested successfully for:
    • fp32
    • fp16
    • int8
    • including INT8 with dynamic shapes on TensorRT 11 ✅

• ⚡ Better quantization workflow for modern NVIDIA deployment

  • FP16 is now applied by baking mixed precision into the ONNX graph before engine build.
  • INT8 is now applied through explicit quantization in the ONNX graph with calibration data.
  • This is especially important because TensorRT 11 removed the old methods many exporters depended on.

• 🎯 Big tracking expansion: 4 new built-in trackers

  • PR #24371 added four new multi-object trackers alongside BoT-SORT and ByteTrack:
    • OC-SORT
    • Deep OC-SORT
    • FastTracker
    • TrackTrack
  • These are now documented and wired into the tracking system with YAML configs like:
    • ocsort.yaml
    • deepocsort.yaml
    • fasttrack.yaml
    • tracktrack.yaml

• 🏃 Tracking docs and selection guidance improved

  • Tracking docs were expanded to explain the strengths of each tracker and how to choose between them.
  • This makes the tracking feature much easier to use for both beginners and advanced users.

• 📹 Video stream loading is safer

  • PR #24749 improves cleanup when stream initialization fails.
  • If one stream fails while others already opened, Ultralytics now properly closes those partially-opened resources instead of leaking threads or capture handles.

• ⚡ AI Gym pose workflow is faster

  • PR #24744 reduces repeated GPU-to-CPU syncs in the workout monitoring loop by transferring keypoints to CPU in one go instead of piece by piece.
  • Same behavior, better efficiency.

• 🧪 Validation mixed precision handling simplified

  • PR #24736 consolidates validation autocast into one cleaner scope during training validation.
  • This helps keep mixed-precision behavior simpler and more robust.

• 📝 Documentation improvements

  • Added new Rust inference documentation for running YOLO models through ONNX Runtime without Python.
  • Expanded tracking docs and reference pages.
  • Updated TensorRT and Jetson docs to explain TensorRT 11 behavior and DLA limitations.
  • Refreshed guides like Coral Edge TPU and semantic image search.

🎯 Purpose & Impact

• 🚀 TensorRT 11 users can export again

  • This is the biggest user-facing change in the release.
  • If you deploy on modern NVIDIA systems using TensorRT 11, exports that were failing should now work again.
  • This is especially valuable for production inference pipelines and edge/server deployment.

• ⚙️ Future-proofs NVIDIA deployment

  • TensorRT 11 changed how precision and quantization are handled.
  • By moving to a ModelOpt-based workflow, Ultralytics stays compatible with newer NVIDIA tooling instead of relying on removed APIs.

• 💾 Smaller, faster engines remain accessible

  • FP16 and INT8 exports are still available even with TensorRT 11’s breaking changes.
  • That means users can continue optimizing for:
    • faster inference
    • lower memory use
    • better deployment efficiency

• 📦 Dynamic INT8 export support is especially useful

  • Supporting INT8 with dynamic shapes on TensorRT 11 can help users deploying across varying input sizes without giving up quantization benefits.

• 👀 Tracking becomes more flexible for real-world scenarios

  • The new trackers give users more choices depending on their needs:
    • simple baseline tracking
    • crowded-scene tracking
    • appearance-aware tracking
    • occlusion-heavy tracking
  • This can improve results in surveillance, sports, traffic, and retail use cases.

• 🔒 More stable long-running applications

  • Stream cleanup improvements reduce the chance of lingering resources in apps that open cameras or network streams.
  • This matters most for production or multi-stream systems.

• ⚡ Small but meaningful performance wins

  • AI Gym and validation updates help reduce overhead and improve runtime efficiency without changing how users interact with the API.

Overall, v8.4.63 is a strong deployment-focused release 📦—with the headline improvement being restored and modernized TensorRT 11 export support, plus a major boost to tracking capabilities and several reliability/performance refinements.

What's Changed

• Consolidate validation autocast scope by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24736
• Convert SAM3 docstrings to Google style by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24741
• Merge isolated_model and isolated_task_model into isolated_model_path by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/24742
• Avoid per-keypoint gpu sync in workout monitoring loop by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/24744
• New TrackTrack, FastTracker, OC-SORT and Deep OC-SORT Trackers by @onuralpszr in https://github.com/ultralytics/ultralytics/pull/24371
• Reduce CI PyTorch index flakiness by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24753
• Fix Codecov OIDC uploads by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24754
• docs: 📝 Add Rust Inference documentation by @onuralpszr in https://github.com/ultralytics/ultralytics/pull/24712
• Report SystemLogger disk list by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24758
• Fix instances count logging by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/24738
• Clean up partially-initialized LoadStreams on failure by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/24749
• Restructure Coral Edge TPU guide and fix FAQ Edge TPU model filename by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/24745
• Restructure semantic image search guide by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/24746
• NVIDIA TensorRT 11 support with ModelOpt FP16 and INT8 quantization by @onuralpszr in https://github.com/ultralytics/ultralytics/pull/24735

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.62...v8.4.63