🌟 Summary

v8.4.172 makes oriented-box cropping work as expected and improves reliability across model training, export, dataset handling, and deployment.

📊 Key Changes

• Rotation-aligned OBB crops: save_crop=True and ObjectCropper now save crops upright and aligned to each oriented bounding box. Areas extending beyond the image are filled with black.
• Model correctness fixes: AIFI positional embeddings now match non-square feature maps, and RT-DETR computes matching costs in float32 to avoid AMP precision errors affecting training assignments.
• More reliable training and resume: NaN recovery now correctly restarts the recovered epoch, while resumed runs preserve the selected optimizer and early-stopping progress.
• Improved dataset and export handling: Corrupted dataset caches trigger a rescan; classification datasets recognize more supported image formats; and static INT8 exports can calibrate even when the dataset is smaller than the requested batch.
• Broader Python support: Python 3.14 is added to package support and several Docker and export environments. The Python export image remains on Python 3.13 for dependencies that do not yet support 3.14.
• Platform documentation refreshed: Docs now describe current Platform behavior and APIs, including Agents management, dataset version comparisons, and direct LabelMe imports.

🎯 Purpose & Impact

• OBB users can now get useful, tightly aligned crops instead of no crops or axis-aligned approximations.
• Training and validation should be more dependable, particularly when using AMP, recovering from NaNs, or resuming an interrupted run.
• Dataset and export workflows are more resilient to damaged caches, varied image formats, and small calibration datasets.
• Updated Python support and clearer Platform documentation make it easier to choose compatible environments and follow current workflows.

What's Changed

• Fix NaN conv outputs in JetPack5 Docker image from OpenBLAS 0.3.8 by @glenn-jocher in #26473
• Bump Prettier from 3.8.5 to 3.9.9 by @glenn-jocher in #26472
• Remove cloudpickle and filelock base dependencies and disallow new ones by @glenn-jocher in #26471
• Remove redundant Lost-state checks in ByteTrack and FastTracker second association by @glenn-jocher in #26481
• Use pre-cached assets in COCO-eval and CLI solutions tests by @JaviChulvi in #26476
• Align validation, Edge TPU export, and Platform dataset, Agents API, and billing docs with current behavior by @raimbekovm in #26467
• Fix AIFI position embedding for non-square feature maps by @ArgoHA in #26483
• Align Ultralytics Platform docs with current Platform behavior by @pderrenger in #26485
• Move AMD doc to Deployment Integrations by @lakshanthad in #26489
• Name LabelMe and Labelbox in Platform import format lists and correct Label Studio OBB and VOC notes by @raimbekovm in #26488
• Use >= floors in example requirements to stop Dependabot pin bumps by @dependabot[bot] in #26501
• Restore the shared CI asset cache in Docker tests and keep it out of images by @glenn-jocher in #26502
• Fix corrupt dataset caches, NaN recovery, YOLOE predict state and batching, RT-DETR save_txt, and tracker/cfg/solution bugs by @cainiao33 in #26477
• Fix RT-DETR Hungarian matching precision under AMP by @Nikhi00718 in #26475
• Restore the optimizer=auto choice and early stopping state on resume by @raimbekovm in #26487
• Replace outputs from previous predict and val runs and give each video its own output name by @raimbekovm in #26494
• Fix integer crop coordinates and JPEG output collisions by @Vaishnavi220506 in #26474
• Upgrade CPU containers and isolated exports to newer Python versions by @glenn-jocher in #26510
• Load every supported image format in classification datasets and refresh stale disk cache by @raimbekovm in #26497
• Fix OpenVINO INT8 calibration for undersized datasets by @JaviChulvi in #26507
• Fix yolo_bbox2segment background-first detection and rerun duplicates, and accept jpeg and uppercase masks by @wizzseen in #26508
• Add rotation-aligned OBB crops to save_crop by @Nikhi00718 in #26486

New Contributors

• @ArgoHA made their first contribution in #26483

Full Changelog: v8.4.171...v8.4.172

🌟 Summary

v8.4.171 adds end-to-end AMD GPU support for Ultralytics workflows, alongside fixes to dataset handling, model export, and several task-specific behaviors.

📊 Key Changes

• AMD ROCm and MIGraphX support (PR #24137): Train and run native PyTorch models on supported AMD GPUs with ROCm, and run exported ONNX models using the MIGraphX execution provider. Provider selection is automatic on supported systems, with CPU fallback if MIGraphX is unavailable.
• AMD deployment options: Adds a ROCm-based latest-amd Docker image, AMD GPU CI testing, and setup and usage guidance. MIGraphX support targets Linux x86_64 with Python 3.11 or newer.
• Faster repeat ONNX startup: MIGraphX compiled models are cached, avoiding recompilation on later loads. Published Radeon 8060S benchmarks show matching accuracy and faster inference for ONNX/MIGraphX than PyTorch ROCm in the tested configurations; the first run still incurs compilation time.
• Dataset and training fixes: Restores reading existing image caches across cache modes, preserves archive timestamps to avoid unnecessary label rescans, and corrects RT-DETR denoising query matching during training.
• Export and task fixes: Reduces temporary storage used by Core ML NMS exports, improves pose keypoint scaling in LiteRT/RKNN, and fixes several mask and analytics behaviors.
• Documentation and CI improvements: Adds Apple M4 Core AI/Core ML latency results and suppresses expected tracer warnings in pytest logs.

🎯 Purpose & Impact

• More hardware choice: AMD users can train with ROCm and deploy ONNX models with GPU acceleration through MIGraphX, without changing their usual prediction calls.
• Smoother deployment: The AMD Docker image and dedicated setup documentation make it easier to get started, while hardware CI helps verify the ROCm path.
• More reliable workflows: Cache and training fixes help avoid wasted disk space, repeated dataset processing, and incorrect RT-DETR denoising matches.
• No new model architecture is introduced: The headline change is broader AMD hardware support, with additional bug fixes and clearer performance guidance.

What's Changed

• Fix regressions found in third-party PR audits by @glenn-jocher in #26461
• Add Core ML and Core AI latency benchmarks on Apple M4 to the Core AI docs by @onuralpszr in #26459
• Remove extra temporary weights copy from CoreML NMS export by @amanharshx in #26466
• Fix RT-DETR contrastive denoising group indices by @Nikhi00718 in #26465
• Ignore benign TracerWarnings in pytest like the package does by @UltralyticsAssistant in #26462
• Add AMD ROCm and MIGraphX support by @itikhono in #24137

New Contributors

• @itikhono made their first contribution in #24137

Full Changelog: v8.4.170...v8.4.171

🌟 Summary

Ultralytics 8.4.170 restores the established disk-cache naming, so existing caches are reused instead of rebuilt. It also documents video frame skipping for Platform inference.

📊 Key Changes

• 💾 Reverted the disk-cache filename change: Detection and classification datasets again save cache files as image.npy, rather than image.jpg.npy.
• 🎥 Documented vid_stride for Platform inference: The parameter controls how often video frames are processed; its default of 1 processes every frame, and images are unaffected.
• 📖 Updated video-response docs: They now clarify that results are returned for each processed frame.

🎯 Purpose & Impact

• ✅ Users with existing cache=disk datasets can reuse their original cache files without rebuilding a second copy, saving time and disk space.
• ⚠️ This restores compatibility but gives up the previous change’s handling of same-named images with different extensions in one folder. For detection datasets, those images also share a label file, limiting the practical benefit of separate cache names.
• 🚀 Platform users can use vid_stride to process fewer video frames when the API supports it, potentially reducing inference work. No model architecture changes are included.

What's Changed

• Avoid disk cache collisions for same-stem images by @aswanth-07 in #26457
• Document vid_stride for Platform predict by @glenn-jocher in #26458
• ultralytics 8.4.170 Revert disk cache naming from #26457 by @glenn-jocher in #26460

Full Changelog: v8.4.169...v8.4.170

🌟 Summary

Ultralytics v8.4.169 improves dataset reliability, preserves more COCO annotations, and fixes Edge TPU export setup when Google’s old package repository is unavailable.

📊 Key Changes

• Unique disk-cache files: Cache filenames now keep the original image extension, so files like sample.jpg and sample.png no longer share a cache file. This prevents one image’s pixels from being mistakenly used for another, in both detection-family and classification datasets.
• COCO conversion preserves distinct annotations: Segmentation polygons and pose keypoints are no longer discarded just because they share a class and bounding box. Exact duplicate output annotations are still removed, and detection continues to deduplicate boxes.
• More reliable Edge TPU export: If the Edge TPU compiler is missing, it is downloaded from an Ultralytics asset instead of Google’s retired apt repository. This avoids apt and sudo requirements; an installed compiler on PATH is still used when available. The bundled compiler supports Linux x86_64.
• Clearer export guidance: Documentation now explains that NCNN, Ultralytics’ MNN backend, and some OpenVINO CPU configurations can run FP32 exports with reduced precision. It also updates guidance on quantization, multi-input exports, and ExecuTorch setup.

🎯 Purpose & Impact

• 🗂️ More trustworthy training data: Disk caching no longer risks silently mixing image contents, including when datasets are loaded in separate subsets.
• 🧩 Fewer lost training labels: COCO segmentation and pose datasets retain distinct annotations needed to train on different shapes or keypoints within the same box.
• 🚀 Fewer export setup failures: Edge TPU exports can install the compiler without relying on Google’s inaccessible apt repository, including in rootless environments.
• 📖 Easier export troubleshooting: Updated documentation helps users understand expected numerical differences between runtimes and choose the correct export options.

What's Changed

• Install the Edge TPU compiler without Google's retired apt repo by @glenn-jocher in #26456
• Name every reduced-precision runtime in the non-YOLO export parity guidance by @raimbekovm in #26455
• Preserve distinct COCO polygons and keypoints sharing a box by @Vaishnavi220506 in #26454
• Avoid disk cache collisions for same-stem images by @aswanth-07 in #26457

Full Changelog: v8.4.168...v8.4.169

🌟 Summary

YOLO 8.4.168 delivers practical fixes for image and mask handling, prediction outputs, video timing, and exported models, alongside clearer training guidance. No model architecture changes are included.

📊 Key Changes

• 🖼️ More reliable image and mask processing: Palette PNG masks now retain their class IDs during conversion, 16-bit semantic masks preserve their IDs and flag invalid classes, and disk caches refresh when source images change.
• 🔄 Correct image orientation: AVIF images with EXIF orientation are read upright.
• 📁 No more overwritten prediction files: Images with the same filename stem, such as bus.jpg and bus.png, get distinct output names.
• 📦 Support for short inputs to fixed-batch exports: Inputs are padded to the required batch size and extra outputs are removed afterward.
• 🎥 Improved video handling: Fractional frame rates are preserved when reading videos, helping avoid playback timing drift.
• 🧩 Semantic-mask CutMix works as intended: CutMix can paste semantic-mask regions even when there are no object instances.
• 🧠 Quieter ReID setup: Re-identification models initialize embeddings with an in-memory image, avoiding a missing-source warning and unnecessary asset-image inference.
• 📚 Clearer training and platform guidance: Documentation now explains that mask_ratio affects training masks, while optimizer=auto ignores lr0; it also clarifies deployment rename behavior. CLI examples were adjusted accordingly.
• 🧪 Lighter similarity-search testing: SearchApp is tested using existing test images rather than triggering an extra large image download.

🎯 Purpose & Impact

• ✅ Reduces common data-handling errors that could cause incorrect labels, stale training images, or unexpected prediction-file overwrites.
• 🚀 Makes inference more dependable across fixed-batch exported models and varied image and video formats.
• 🎯 Helps semantic-segmentation users get the intended CutMix augmentation.
• 💡 Sets clearer expectations for training settings, so users are less likely to pass options such as lr0 that have no effect with the default automatic optimizer.

What's Changed

• Clarify mask_ratio does not change predicted mask resolution by @Y-T-G in #26441
• Warm the ReID predictor with an in-memory image by @raimbekovm in #26453
• Fold SearchApp init into the similarity search test to avoid a 32 MB download by @raimbekovm in #26452
• Correct optimizer=auto lr0/momentum docs and document deployment rename slug by @raimbekovm in #26447
• Fix CutMix skipping every paste on PNG semantic mask datasets by @Nikhi00718 in #26439
• Fix 16-bit semantic mask ids, CLI run name parsing, and fractional video fps by @Nicholas022400701 in #26440
• ultralytics 8.4.168 Fix palette mask conversion, stale disk cache, AVIF orientation, same-stem predict outputs, and static-batch short inputs by @cainiao33 in #26443

Full Changelog: v8.4.167...v8.4.168

🌟 Summary

Ultralytics 8.4.167 reduces repeated downloads and storage for Platform dataset versions by reusing unchanged, content-addressed assets. No model architecture or training behavior changes are included.

📊 Key Changes

• ♻️ Reuses Platform dataset assets: Images and depth files with matching content-addressed URLs can be hard-linked from a shared cache, so new dataset versions don’t need to download duplicate files.
• 🛡️ Protects shared files: Downloads are written to temporary files and moved into place only when complete. Image repair and compression also replace files safely, avoiding unintended changes to assets shared between versions.
• 🔄 Preserves compatibility: Pooling is limited to eligible Ultralytics Platform manifests. Other data sources, and filesystems that cannot create hard links, continue using the existing download behavior.
• 📚 Refreshes documentation: Updated tutorial videos and captions, and added references for the new image utility helpers.
• ⚙️ Updates CI maintenance: Bumped the self-hosted runner cleanup action to v1.4.41.

🎯 Purpose & Impact

• 💾 Less storage and network use: Unchanged assets can be reused across dataset versions instead of stored and downloaded again—especially helpful for large datasets and persistent Platform workers.
• ✅ Safer dataset processing: Atomic file replacement helps prevent incomplete downloads and protects shared assets during image cleanup or compression.
• 👥 No workflow change for most users: The optimization is automatic for supported Platform datasets; other manifests retain their previous behavior.

What's Changed

• Bump eviden-actions/clean-self-hosted-runner from v1.4.40 to v1.4.41 in /.github/workflows by @UltralyticsAssistant in #26448
• Add https://youtu.be/GW5z7HX4Jds & https://youtu.be/D3W9SuEWDOk to docs by @RizwanMunawar in #26425
• ultralytics 8.4.167 Share content-addressed NDJSON assets across dataset versions by @glenn-jocher in #26442

Full Changelog: v8.4.166...v8.4.167

🌟 Summary

Ultralytics 8.4.166 focuses on more reliable dataset handling and prediction outputs, alongside updated Platform and inference documentation. No model architecture changes are included.

📊 Key Changes

• 🗂️ Semantic masks and dataset conversion: The release preserves class IDs in Pascal VOC-style palette PNG masks and retains pixel values from 16-bit masks. It also rebuilds older semantic-mask caches and includes LVIS images even when they have no annotations.
• 🎯 More accurate segmentation outputs: Fixes prevent resized instance masks from shrinking during validation and preserve the correct letterbox crop when saving semantic predictions.
• 🧰 Dataset and training reliability: Fixes address stale disk-cached images, relative directories and globs in prediction source lists, outdated classification split copies, and unsafe small-image training configurations.
• 🖼️ Image and video handling: Annotation visualization and image compression now account for EXIF orientation, and saved videos retain the expected duration when frames are skipped.
• 📚 Documentation updates: Inference guidance now reflects Rust inference 0.0.50, including device selection, video requirements, GPU preprocessing, and RT-DETR support. Platform docs also describe current annotation, dataset, inference, and deployment behavior.

🎯 Purpose & Impact

• ✅ Better compatibility with real-world datasets: Palette-based and 16-bit masks are less likely to be rejected or interpreted with incorrect class IDs.
• 📐 More trustworthy segmentation results: Validation masks and saved predictions better match the original image dimensions and content.
• 🛠️ Fewer confusing failures: Dataset loading, training checks, and prediction inputs behave more consistently, catching some issues earlier.
• 📖 Clearer guidance for users: Updated docs help users understand current Platform features and inference defaults, including the Ultralytics Platform.

What's Changed

• Docs: document Platform image generation, similarity, and face blurring by @JaviChulvi in #26411
• Align Rust inference docs with ultralytics-inference 0.0.50 by @raimbekovm in #26431
• Sync platform docs and screenshots with annotation, inference, and deployment changes by @raimbekovm in #26419
• Classify unsupported onnx opset rejections as expected in fuzzing by @raimbekovm in #26430
• Include the multi_scale minimum size in the final batch=1 training check by @raimbekovm in #26429
• Clarify optimizer=auto learning rate behavior by @zkoymen in #26435
• Respect cache mode when loading existing .npy images by @aswanth-07 in #26432
• Resolve relative directories and globs in prediction source lists by @Vaishnavi220506 in #26420
• Fix semantic save_json letterbox crop, yolo_bbox2segment on NumPy>=2.4, and stale classify re-split copies by @cainiao33 in #26423
• Threshold rescaled segment val masks and keep saved video duration with vid_stride by @Nicholas022400701 in #26416
• Apply EXIF orientation in image utilities and fix TQDM ':60' time rollover by @MohammadHijjawi97 in #26414
• ultralytics 8.4.166 Support Pascal VOC-style palette and 16-bit PNG masks by @Nikhi00718 in #26426

New Contributors

• @zkoymen made their first contribution in #26435

Full Changelog: v8.4.165...v8.4.166

🌟 Summary

v8.4.165 improves package compatibility, export and inference reliability, and dataset handling, with several targeted fixes for model outputs and vision workflows.

📊 Key Changes

• 📦 Cleaner Python wheel installation (current PR #26408): Test files are no longer installed as a top-level tests package, avoiding conflicts with other projects. Tests remain in the source distribution so Conda builds can still run them.
• 🚀 More reliable exports and inference: ONNX export now checks that the requested opset is high enough for the model and reports a clear error if it is not. Windows OpenVINO inference also avoids a CPU instruction path that could crash on some machines.
• 📐 Improved depth and semantic outputs: Depth resizing now follows the same alignment convention used during training and validation. Saved semantic prediction masks are restored to the original image size without letterbox padding, and invalid semantic class IDs are flagged during dataset scanning.
• 🧹 Safer, more consistent dataset processing: Fixes cover whitespace-only labels, distinct segmentation polygons that share a bounding box, repeatable classification dataset splits, older Albumentations versions, and nested paths in COCO/LVIS conversion. Multispectral conversion now keeps source images if outputs collide or fail to write.
• 🎯 More accurate vision utilities: Corrected flips for left-top-width-height boxes, kept tracked pose keypoints colored by track ID, improved distance-calculation selection, and fixed class histories in analytics charts.
• 🛠️ Documentation and CI updates: Clarified Platform annotation and image-search guidance, corrected Ascend device mappings, and updated runner versions and test behavior.

🎯 Purpose & Impact

• ✅ Fewer installation surprises: The wheel no longer shadows other packages’ tests imports, while source-based Conda builds retain the tests they need.
• 🧪 Clearer failure feedback and more dependable deployment: Unsupported ONNX opsets fail early with an actionable message, and Windows OpenVINO is less prone to native crashes.
• 📊 More trustworthy training and evaluation: Dataset issues are caught earlier, repeated classification splits stay consistent, and saved semantic masks and depth outputs better match their intended image geometry.
• 🛡️ Reduced risk of losing data: Failed or colliding multispectral conversions no longer silently delete original images.

What's Changed

• Fix Windows OpenVINO crashes and CI runtime warnings by @glenn-jocher in #26317
• Update GitHub Actions runner to 2.337.0 in runner images by @glenn-jocher in #26375
• Refresh AGENTS.md: stale docformatter claim and workflow guidance by @glenn-jocher in #26381
• Treat unreadable paths as missing in check_file by @glenn-jocher in #26382
• Docs: fix swapped Ascend310B1/B4 device mapping in Ascend export table by @Tsingxuan in #26394
• Fix custom spatial Albumentations lists on detection labels with older albumentations releases by @raimbekovm in #26399
• Fix flips of left-top-width-height instance boxes by @Nikhi00718 in #26397
• Cap oneDNN below AMX for Windows OpenVINO inference by @glenn-jocher in #26403
• Fail fast on xdist worker crashes in CI by @glenn-jocher in #26405
• Color tracked keypoints by track ID in Results.plot instance mode by @MohammadHijjawi97 in #26389
• Update Platform docs for annotator panning, blur faces, and similar image search by @raimbekovm in #26401
• Select one object per click in DistanceCalculation by @MohammadHijjawi97 in #26402
• Skip whitespace-only lines when parsing label files by @cainiao33 in #26391
• Use align_corners=True for depth export and predict resize by @cainiao33 in #26393
• Name the grounding data format for from-scratch World trainers and document the empty background class by @raimbekovm in #26398
• Keep classification auto-split deterministic so re-runs do not leak val into train by @MohammadHijjawi97 in #26387
• Fix Analytics area chart mixing up class histories by @MohammadHijjawi97 in #26386
• Create label subfolders in convert_coco when file_name has a directory by @MohammadHijjawi97 in #26388
• Keep originals when multispectral conversion collides or fails to write by @Vaishnavi220506 in #26406
• Write LVIS image lists relative to the list file in convert_coco by @MohammadHijjawi97 in #26390
• End the scale_masks crop at the letterbox content when the padding is odd by @Nicholas022400701 in #26379
• Write the semantic val PNG masks at the original image size without the letterbox padding by @Nicholas022400701 in #26385
• Ignore empty paths in TXT and CSV prediction sources by @aswanth-07 in #26371
• Compare the polygons, not their boxes, when dropping duplicate segment labels by @Nicholas022400701 in #26377
• Check minimum ONNX opset before export by @onuralpszr in #26396
• Reject semantic masks with unmapped class ids at scan by @cainiao33 in #26392
• Stop installing top-level tests package from the wheel by @Y-T-G in #26408

New Contributors

• @Tsingxuan made their first contribution in #26394

Full Changelog: v8.4.164...v8.4.165

🌟 Summary

v8.4.164 makes FLOPs reporting more efficient and improves export, video inference, dataset validation, and reliability—without changing model architectures or predictions.

📊 Key Changes

• Faster FLOPs counting (PR #26373): Delegates FLOPs estimation to THOP 2.2.0, which supports attention operations and profiles at smaller strides. This replaces full-size profiling for attention-based models such as YOLO11, YOLO12, and YOLO26, as well as RT-DETR. The PR depends on ultralytics-thop 2.2.0.
• More efficient exports: CoreML export skips redundant trace checking, with reported trace times around 2.8–3× faster. The exporter also removes a duplicate setup forward pass.
• Improved video prediction: CUDA prediction can prefetch video frames while the GPU processes the current batch. Batches now stay within one video, also fixing an FPS mix-up when processing videos with different frame rates.
• More reliable data handling: Dataset scans now catch malformed OBB labels and fractional class IDs earlier. They also accept pose datasets for detection training and correctly parse multi-row prediction CSV files.
• Bug fixes and stability: Fixes include memory growth in PIL-based annotation, cleanup when prediction streams end early, tracker callbacks affecting plain prediction, and YOLOv9 distillation across different head layouts.
• Clearer deployment guidance: Documentation adds and updates guidance for Core AI, Rust inference options, export formats, datasets, and deployment workflows.

🎯 Purpose & Impact

• Less overhead when checking model information or starting training: FLOPs reporting no longer needs a full-resolution forward pass for the affected models, which can make model.info() and training startup logging lighter.
• Faster and smoother deployment workflows: Export setup does less redundant work, and CUDA video prediction can overlap frame loading with GPU inference.
• Earlier, clearer feedback on dataset problems: Invalid labels are more likely to produce helpful errors during scanning rather than confusing failures later in training or validation.
• More dependable runs: Fixes reduce resource leaks, incorrect tracker behavior, and failures in distillation and prediction-source handling.

What's Changed

• Ride out PyPI outages when publishing and alert Slack once by @glenn-jocher in #26340
• Speed up CoreML export by disabling redundant trace checking by @amanharshx in #26337
• Update Platform docs to 21 export formats by @glenn-jocher in #26344
• Update ultralytics-inference version to 0.0.48 and add quantize in documentation by @onuralpszr in #26345
• Fix memory growth in Annotator(pil=True) from a font reference cycle by @Y-T-G in #26346
• Return early from segment2box for segments fully outside the image by @raimbekovm in #26331
• Document Core AI export on Linux and Platform, and 409 on duplicate resource names by @raimbekovm in #26343
• Reject box-only labels for OBB datasets at load time like segment datasets by @Nicholas022400701 in #26330
• Run the exporter dry-run forward pass once by @amanharshx in #26334
• Keep the checkpoint dataset on resume and omit host paths from the export validate hint by @raimbekovm in #26335
• Prefetch video frames during CUDA prediction by @JESUSROYETH in #26349
• Restore Conv.default_act after parse_model overrides it by @cainiao33 in #26347
• Release prediction writers and sources when stream runs end early by @cainiao33 in #26348
• Fix YOLOv9 distillation across teacher and student head layouts by @Nikhi00718 in #26338
• Fix multi-row CSV prediction sources by @Nikhi00718 in #26360
• Fix tracker callbacks running during plain prediction by @aswanth-07 in #26352
• Drop the keypoints in verify_image_label so a detect task can train on a pose dataset by @Nicholas022400701 in #26357
• Isolate Jetson CI settings per runner by @glenn-jocher in #26366
• Apply format-specific export options for mixed-case and alias formats by @MohammadHijjawi97 in #26358
• Join bracketed CLI values after '=' merging in merge_equals_args by @MohammadHijjawi97 in #26365
• Route tar-family dataset download URLs through safe_download by @cainiao33 in #26361
• Remove unused labels autolabel path from non_max_suppression by @cainiao33 in #26356
• Fix ObjectCropper crash with OBB models by @MohammadHijjawi97 in #26359
• 86% faster SAM batched_mask_to_box on CPU by @Y-T-G in #26364
• Fix valid .jp2 images dropped as corrupt in dataset scans by @cainiao33 in #26355
• Reduce Jetson CI time and cancel superseded hardware runs by @glenn-jocher in #26369
• Re-derive teacher feature indices when unpickling distillation checkpoints by @cainiao33 in #26362
• Reject fractional class IDs during dataset scans by @Nikhi00718 in #26367
• Fix docstrings, export docs/args, and bugs found in a full package audit by @glenn-jocher in #26370
• Fix stale TensorRT INT8 calibration cache reuse by @amanharshx in #26368
• Add lite, openapi, and skills to CI status table by @glenn-jocher in #26372
• Auto-select idle GPUs at DDP launch in test_train by @glenn-jocher in #26374
• Delegate FLOPs counting to THOP 2.2.0 stride profiling by @glenn-jocher in #26373

Full Changelog: v8.4.163...v8.4.164

🌟 Summary

YOLO models can now be exported to Apple Core AI format from x86_64 Linux, making Linux servers and CI environments useful for preparing models for Apple devices.

📊 Key Changes

• 🍎 Core AI export on Linux: Export .aimodel files on x86_64 Linux as well as Apple silicon Macs running macOS 26 or later. The exported models target iOS 27 and macOS 27; running them still requires Apple hardware.
• 🔧 Updated Core AI compatibility: Supports coreai-torch 0.4.3, which performs optimization during conversion. The dependency now supports Python 3.11–3.14.
• 📈 Clearer benchmarks: Linux can report Core AI export results, while inference is correctly skipped because it requires macOS.
• 🛠️ More reliable exports: PyTorch functions patched for certain ONNX exports are now restored even when an export fails, preventing unexpected effects on later operations in the same process.
• 🐳 Lean­er Docker images: Docker builds avoid retaining package-install caches and remove bundled PyTorch test files, reducing image contents without removing needed runtime tools.

🎯 Purpose & Impact

• 🌐 Teams can build Apple-targeted models in Linux-based cloud, server, and CI workflows instead of needing a Mac for export.
• 📱 Apple hardware is still required to run .aimodel files, and Core ML remains a separate option.
• ✅ Export failures are less likely to leave a Python process in a modified state, while smaller Docker images use less storage and are quicker to distribute.

What's Changed

• Skip the uv cache and torch test files in Docker images, and fix Core AI export by @glenn-jocher in #26326
• Support coreai-torch 0.4.3, which removed AIProgram.optimize() by @cdeil in #26327
• Restore PyTorch functions after failed exports by @Nikhi00718 in #26325
• ultralytics 8.4.163 Export Apple Core AI models on Linux by @glenn-jocher in #26333

Full Changelog: v8.4.162...v8.4.163

🌟 Summary

YOLO prediction can now load the next batch of images while CUDA processes the current one, alongside performance, export, dataset, and Ultralytics Platform improvements. No model architectures or weights changed.

📊 Key Changes

• 🚀 Faster CUDA prediction: For multi-batch, image-only detection using PyTorch weights, prediction prefetches the next batch on a worker thread while the GPU handles the current batch. Other devices, tasks, backends, and mixed image/video sources keep the existing behavior.
• ⚡ Faster training data preparation: Optimized copy-paste augmentation and semantic-mask target generation while preserving their expected outputs.
• 🧠 Lower memory use during TensorFlow INT8 export: Calibration images are collected into a preallocated array instead of building several large intermediate copies.
• ☁️ More Platform annotation and dataset features: Added hosted text-prompted SAM 3 and SAM 3.1 options, expanded cloud-storage imports to 500,000 objects, and enabled image search by annotated class name.
• 🛠️ More robust workflows: Fixed Windows OpenVINO crashes and GPU ONNX inference silently falling back to CPU; improved dataset archive and .yml support, OOM recovery, and tracking when source files share a name.

🎯 Purpose & Impact

• ⏱️ Better GPU utilization: Overlapping image loading with inference can reduce idle time during eligible CUDA prediction runs. It complements existing work that speeds up decoding within a batch.
• 📈 Smoother data workflows: Faster augmentation and reduced calibration memory use can help improve training and export efficiency, especially on constrained systems.
• 🔍 Easier annotation and dataset management: Text-prompted SAM, larger cloud imports, and class-name search make it simpler to label and browse larger datasets.
• ✅ Fewer surprises: Export, Windows inference, archive loading, tracking, and training recovery fixes improve reliability without changing model behavior.

What's Changed

• Reduce peak memory during INT8 TensorFlow calibration by @cainiao33 in #26288
• Point the validator at the rebuilt dataloader after OOM auto-reduce by @cainiao33 in #26296
• Simplify tar download handling and remove redundant regression fixtures by @cainiao33 in #26287
• fix incorrect dataset in OBB loss error message by @lmycross in #26314
• Fix Windows OpenVINO crashes and CI runtime warnings by @glenn-jocher in #26317
• Remove headless-unsafe destroyAllWindows from Streamlit Inference by @glenn-jocher in #26320
• Annotate CI test failures and xdist worker crashes on GitHub Actions by @Y-T-G in #26318
• Speed up overlap_mask=False semantic mask targets with a single min reduction by @raimbekovm in #26315
• Speed up CopyPaste with a masked cv2.copyTo by @raimbekovm in #26310
• Discover .yml dataset YAMLs in directories and archives by @Nikhi00718 in #26323
• Accept list and tuple boxes in xyxy2xywh and xywh2xyxy by @MohammadHijjawi97 in #26322
• Reset the tracker between sources that share a filename by @aswanth-07 in #26312
• Document text-prompted SAM, 500K-object cloud imports, endpoint warm-up states and class-name image search by @raimbekovm in #26311
• Extract .tar.xz, .tar.bz2 and local .tgz dataset archives by @Vaishnavi220506 in #26306
• Return None for empty image files in imread_unicode by @cainiao33 in #26304
• Document segments2boxes inputs as (N, 2) point arrays by @MohammadHijjawi97 in #26321
• Close the replaced train dataloader on OOM auto-reduce by @cainiao33 in #26302
• ultralytics 8.4.162 Prefetch the next image batch during CUDA prediction by @JESUSROYETH in #26319

Full Changelog: v8.4.161...v8.4.162

🌟 Summary

Ultralytics 8.4.161 is a reliability and usability release: it fixes several training, inference, plotting, and upload bugs, while improving dataset handling and clarifying export and Platform documentation.

📊 Key Changes

• 🛠️ Fixed six package issues in PR #26301, the release’s main change:
  • RT-DETR decoder tensors now move correctly between devices, helping prevent validation or prediction device-mismatch errors.
  • Resuming older Adam/AdamW checkpoints now preserves the optimizer’s runtime fused setting.
  • Empty image files no longer crash prediction runs.
  • Training results can be read from paths containing bracket characters, and plots handle odd numbers of result panels correctly.
  • Platform uploads can fall back to last.pt when save=False means no best.pt was created.
• 🗂️ Improved dataset and training utilities: .tgz archives now extract correctly, semantic segmentation can skip unnecessary polygon processing, and fuzz testing better recognizes expected missing-split errors.
• 🍎 Made CoreML export failures clearer: removed an invalid fallback that could hide the original save error behind a misleading file-extension error.
• 📘 Expanded export guidance: documented the device option for Hailo and Ascend, and explained why static exports may produce different predictions on non-square images.
• ☁️ Refreshed Platform documentation for dataset image copying and face blurring, annotation providers, deployment pages and monitoring, and related API and billing details.

🎯 Purpose & Impact

• ✅ More dependable prediction and training-resume workflows, especially with RT-DETR, older checkpoints, and imperfect image datasets.
• 📈 Fewer failures when reading results or generating plots, including with unusual file paths and uneven metric layouts.
• ⚡ Less unnecessary processing for semantic datasets where polygon data is not used.
• 🔍 More actionable CoreML errors and clearer export settings help users diagnose problems and choose deployment options.
• 🧭 Updated Platform guidance makes dataset, annotation, billing, and deployment workflows easier to understand.

What's Changed

• Document device export argument for OpenVINO, Hailo and Ascend by @raimbekovm in #26286
• Tighten export mode docs and explain static export prediction differences by @Vaishnavi220506 in #26285
• Align Platform docs with deployment pages, image copy, face blurring and new annotation providers by @raimbekovm in #26297
• Classify wrapped dataset split errors as expected in CLI fuzzing by @aswanth-07 in #26299
• Remove invalid CoreML .mlmodel save fallback by @amanharshx in #26298
• Fix extraction of .tgz dataset archives by @aswanth-07 in #26294
• Skip polygon resampling for semantic datasets when no augmentation reads the polygons by @raimbekovm in #26295
• Fix RT-DETR device moves, fused resume, empty images, bracket paths, odd plot grids and save=False uploads by @glenn-jocher in #26301

New Contributors

• @Vaishnavi220506 made their first contribution in #26285

Full Changelog: v8.4.160...v8.4.161

🌟 Summary

Ultralytics 8.4.160 makes dataset edits less likely to go unnoticed and improves memory use, training reliability, and everyday API behavior—without introducing a new model architecture.

📊 Key Changes

• Dataset cache invalidation (PR #26283): Cache keys now account for each file’s size and modification time, so editing a label in place—such as changing its class without changing its length—triggers a rescan. Detection, depth, semantic, and other shared-cache users benefit.
• Lower peak VRAM for SAM3 text prompting: Text prompts are processed in chunks of up to 16, helping reduce memory use on large prompt sets and avoid out-of-memory errors.
• More reliable training after out-of-memory errors: When training retries with a smaller batch size, it now preserves optimizer and scheduler state instead of restarting optimizer state.
• Clearer validation dataset selection: Validation uses the checkpoint’s recorded dataset when it can be found; otherwise, it warns and falls back to the task’s default dataset. Users can specify data to choose a dataset explicitly. Depth calibration now requires data.
• Lower host RAM during TensorRT 11 INT8 calibration: Calibration releases consumed batch data and normalizes in place, reducing peak host memory without changing calibration values.
• Safer and more convenient utilities: CUDA-backed results can now convert to NumPy; NMS no longer modifies raw prediction tensors; and box-cropping, image-to-label path conversion, and repeated downloads of single-file ZIPs handle common inputs more reliably.
• Documentation improvements: Task guides clarify validation behavior, result fields, dataset formats, and training options, including cloud training.

🎯 Purpose & Impact

• More trustworthy dataset updates: Changed labels are less likely to remain silently stale in cached datasets. Existing caches will rescan once after upgrading; if a file’s size and modification time are both preserved, remove the affected cache manually.
• Better usability across hardware and workflows: SAM3 can handle more text prompts within available GPU memory, while TensorRT 11 calibration can use less host RAM.
• More dependable training and inference: OOM retries retain learned optimizer state, and repeated NMS calls no longer risk corrupting the original predictions.
• Fewer surprising API failures: CUDA results and common Python path or box inputs work more consistently.
• Less ambiguity in validation: Checkpoint dataset settings are reused when available, with a warning and sensible fallback when they are not.

What's Changed

• Add https://youtu.be/VxP9kHUwE6o to docs by @RizwanMunawar in #26274
• Reduce peak VRAM for SAM3 text prompting via chunked grounding by @Laughing-q in #26281
• Validate on the checkpoint's dataset when val gets no data and require data= in calibrate() by @raimbekovm in #26277
• Reduce host memory during TensorRT 11 INT8 calibration by @amanharshx in #26275
• Fix Results.numpy() for CUDA-backed results by @aswanth-07 in #26282
• Fix repeated safe_download of single-file ZIP archives by @Nikhi00718 in #26273
• Stop non_max_suppression() from mutating input predictions by @Nikhi00718 in #26278
• Fix save_one_box() for list, tuple and NumPy boxes by @MohammadHijjawi97 in #26279
• Accept pathlib.Path inputs in img2label_paths() by @MohammadHijjawi97 in #26280
• Preserve optimizer state when OOM auto-reduces batch size by @Nikhi00718 in #26284
• Unify task docs structure and align val and Results guidance by @raimbekovm in #26271
• ultralytics 8.4.160 Invalidate dataset caches when files are edited in place by @aswanth-07 in #26283

New Contributors

• @MohammadHijjawi97 made their first contribution in #26279

Full Changelog: v8.4.159...v8.4.160

🌟 Summary

v8.4.159 improves dataset split handling, startup monitoring, and INT8 SavedModel export efficiency.

📊 Key Changes

• 📥 Avoids unnecessary dataset downloads during training

  • NDJSON-to-YOLO conversion now respects the requested training validation split.
  • By default, training with split=val skips unused test images, reducing download time, storage, and preparation overhead.
  • Using split=test retains test images and makes the trainer validate against that selected split.
  • Split-aware caching prevents validation and test conversions from incorrectly reusing the same cached dataset.
  • Behavior remains compatible with standalone conversion, metadata inference, classification datasets without validation data, and depth-calibration provenance.
  • YOLO, RT-DETR, and classification training now consistently use the selected validation split and corresponding dataset fractions.

• 📊 Reports system metrics as soon as training starts

  • The existing training_started Platform request now includes an initial system snapshot.
  • CPU, memory, disk, and related monitoring data can appear before the first epoch completes.
  • Epoch-level reporting remains unchanged, with no additional requests, timers, or periodic monitoring signals.
  • Optional system monitoring remains isolated so it does not become a required training dependency.

• 🧠 Reduces peak memory during INT8 SavedModel export

  • Calibration data is now created only when needed for the ONNX-to-SavedModel conversion and released immediately afterward.
  • Testing showed an approximately 533.7 MiB, or 18.2%, reduction in peak process memory on an Apple M4 system.
  • Export behavior, runtime, calibration data, and tested full-integer TFLite outputs remain unchanged.

• 🧪 Adds and updates tests and documentation

  • Tests now cover validation/test split selection, cache behavior, and depth calibration handling.
  • The Platform callback documentation now includes the internal system-metrics helper.

🎯 Purpose & Impact

• ⚡ Faster, lighter dataset preparation: Training workflows using Platform NDJSON datasets should avoid downloading images that are not needed for the selected split.
• ✅ More predictable validation: The split setting now consistently controls which data is prepared, loaded, and evaluated across supported tasks.
• 👀 Better training visibility: Platform users can see system resource information immediately when a run begins instead of waiting for the first epoch.
• 💻 More reliable exports on memory-constrained devices: INT8 SavedModel exports require less peak memory, reducing the risk of out-of-memory failures while preserving existing export results.
• 🔄 Minimal workflow disruption: Existing epoch monitoring, standalone conversion, classification fallbacks, and export behavior are preserved.

What's Changed

• Reduce peak memory during INT8 SavedModel export by @amanharshx in #26272
• Skip unused test downloads and report initial training system metrics by @glenn-jocher in #26276

Full Changelog: v8.4.158...v8.4.159

🌟 Summary

Ultralytics 8.4.158 improves YOLO26 training reliability, model export compatibility, data handling, inference backends, and Ultralytics Platform workflows. 🚀

📊 Key Changes

• 🔄 Mosaic augmentation now stays closed after OOM recovery

  • Fixed a training issue where automatic batch-size reduction rebuilt the dataloader and unintentionally re-enabled Mosaic augmentation.
  • The trainer now re-arms the mosaic-closing logic after rebuilding the pipeline, ensuring the augmentation state matches the training log and configuration.

• ⚙️ More reliable AutoBatch sizing

  • AutoBatch no longer treats a failed batch-one probe caused by BatchNorm at very small image sizes as a GPU memory limit.
  • Added CUDA coverage for small image sizes to prevent regressions.

• 🧠 Improved data loading and caching

  • RAM image caches can now be shared more efficiently with spawn and forkserver DataLoader workers.
  • Linux DataLoader workers start lazily, reducing the risk of image-decoder deadlocks during initialization.
  • Dataset hyperparameters are copied before transforms modify them, preventing one dataset from changing settings used by later datasets.
  • Disk caching now uses a smaller safety margin and handles failed cache writes more safely.

• 📦 Stronger download and dataset handling

  • safe_download() can resume interrupted downloads with HTTP Range requests, handle encoded responses, and better tolerate temporary server errors.
  • Invalid files named as archives are now returned as files instead of being mistaken for extracted datasets.
  • NDJSON conversion keeps at least one image when a valid nonzero fraction would otherwise select none.

• 📤 More accurate model exports

  • Exported models with missing or invalid class names now infer the correct class count from the model head instead of defaulting to 999 names.
  • This improves metadata consistency for legacy detection and classification checkpoints.
  • TorchScript inference is stabilized on older PyTorch versions by avoiding a repeated-inference crash.
  • TensorRT now raises clear errors when dynamic shapes are rejected or inference execution fails.

• 🎭 Better SAM2 video masks

  • SAM2 video predictions now correctly enforce non-overlapping masks before thresholding.
  • The highest-scoring object owns each pixel, preventing tracked objects from incorrectly covering one another.
  • CPU semantic segmentation postprocessing is also faster through improved class-index selection.

• 🍎 Expanded Apple deployment support

  • Core AI assets now include model descriptions and are documented as an opt-in path in the Ultralytics iOS SDK and Flutter plugin.
  • Core ML remains the default and recommended Apple deployment format, especially for broader device compatibility.
  • Core AI remains limited to newer Apple operating systems and Apple silicon export environments.

• 🛠️ Ultralytics Platform improvements

  • Added class-prompted annotation using hosted open-source models and paid vision providers for detection datasets with one to 100 classes.
  • Agents workflows now support HTTPS webhook actions in addition to Slack notifications.
  • TIFF originals can be stored without re-encoding.
  • Platform documentation now covers provider API keys, annotation behavior, upload handling, and the corrected ARKitScenes raw-data workflow.

• 📚 Documentation and compatibility updates

  • ExecuTorch guidance now explains that Ultralytics installs a compatible version automatically.
  • ARKitScenes documentation now uses the raw subset and its approximately two-hertz frame sampling process.
  • System font lookup rescans installed fonts when Matplotlib’s cached list is outdated.

🎯 Purpose & Impact

• More predictable training: Mosaic augmentation remains disabled when expected, even after automatic OOM recovery, improving reproducibility and avoiding misleading training behavior.
• Fewer training failures: AutoBatch, caching, DataLoader startup, and dataset conversion fixes make training more robust across hardware, operating systems, and small image sizes.
• Safer deployment: Exported metadata now matches the actual model, while TorchScript and TensorRT failures produce clearer and earlier errors.
• Cleaner segmentation results: SAM2 video users should see fewer overlapping tracked masks and more consistent object ownership.
• Broader platform workflows: Users can annotate datasets with natural-language or vision-language models, connect workflows to web services, and preserve TIFF source files.
• Better Apple integration: Core AI is usable as an opt-in for supported iOS 27 devices, while Core ML remains the safest default for most Apple deployments. 🍎

What's Changed

• Ignore AutoBatch probe failures below the first successful size by @fcakyon in #26224
• Cover AutoBatch at imgsz below 2x stride in the GPU tests by @glenn-jocher in #26248
• Set the Core AI asset description and document Core AI as an SDK opt-in by @glenn-jocher in #26230
• Remove Core AI from the Platform export format docs by @glenn-jocher in #26249
• Resume safe_download retries with Range requests by @glenn-jocher in #26250
• Avoid fork-inherited image decoder deadlocks in Linux data loaders by @glenn-jocher in #26085
• Share RAM image caches across DataLoader workers by @glenn-jocher in #26086
• Make safe_download robust to encoded responses, curl resumes, and 5xx bursts by @glenn-jocher in #26251
• Add class names fallback for legacy checkpoints on export by @venu-banaras in #24715
• Reduce the disk cache safety margin and guard failed cache writes by @Swish78 in #26112
• Remove the ExecuTorch upgrade advice that installs a runtime the package rejects on torch < 2.13 by @raimbekovm in #26263
• Fix repeated TorchScript inference crash on torch<2.1 by @glenn-jocher in #26264
• Apply the SAM2 video non-overlapping mask constraint on the float masks and select the mask owner with max indices by @raimbekovm in #26262
• Select semantic class indices with max instead of argmax on CPU by @raimbekovm in #26261
• Document the ARKitScenes raw subset and its 2 Hz frame sampling by @Bovey0809 in #26260
• Align Platform docs with tiff passthrough, agents webhooks and class-prompted annotation models by @raimbekovm in #26258
• Rescan system fonts in check_font when the cached font list misses by @cainiao33 in #26256
• Raise on rejected TensorRT shapes and executions by @aswanth-07 in #26268
• Return the file itself when an archive-suffixed download is not a zip or tar by @cainiao33 in #26255
• Keep a nonzero ndjson conversion fraction from selecting zero images by @cainiao33 in #26254
• Stop dataset construction from mutating the caller's hyp namespace by @cainiao33 in #26266
• Preserve model class counts in export name fallbacks by @aswanth-07 in #26253
• ultralytics 8.4.158 Keep mosaic closed after an OOM batch auto-reduction by @rahultechenable in #26257

New Contributors

• @Swish78 made their first contribution in #26112
• @venu-banaras made their first contribution in #24715
• @rahultechenable made their first contribution in #26257

Full Changelog: v8.4.157...v8.4.158

🌟 Summary

Ultralytics v8.4.157 makes TensorRT inference faster—up to 20% for FP16 and INT8 engines—while improving YOLOE prompt-free support, Apple Silicon performance, validation reliability, and training stability. 🚀

📊 Key Changes

• ⚡ Faster TensorRT engines (PR #26223, @Y-T-G)

  • Calibrates FP16 conversion with a real image instead of random noise, allowing more layers to safely run in FP16.
  • Rewrites SiLU activations into a TensorRT-fusable form, reducing memory-bound activation kernels.
  • Adds CUDA Graph replay for supported static TensorRT engines, reducing per-inference launch overhead.
  • Expected impact: up to 20% faster TensorRT FP16 and INT8 inference, especially on static-shape engines. CUDA Graph acceleration is not used for dynamic engines, DLA execution, or engines with embedded NMS.

• 🎯 Expanded YOLOE-26 prompt-free inference (PR #26201)

  • Supports checkpoints containing both one-to-many and one-to-one detection heads.
  • The nms setting can now select between standard NMS inference and NMS-free inference.
  • set_vocab() can regenerate both branches for custom prompt-free YOLOE models.

• 🍎 Improved Apple Silicon CPU and MPS performance

  • Disables a slow NNPACK convolution path on Apple Silicon CPUs, improving larger-batch training and validation.
  • Reworks MPS box-IoU and task-assignment operations to avoid expensive reductions.
  • Replaces semantic segmentation bincount operations with MPS-friendly alternatives, addressing severe slowdowns and buffer-size failures.
  • Optimizes result export, plotting, and summaries by avoiding repeated per-detection conversions.

• 🧪 More reliable training and validation

  • AutoBatch no longer treats an initial batch-size probe failure as an out-of-memory limit.
  • Mosaic augmentation now closes correctly during very short training runs.
  • Resume errors now report the actual checkpoint or configuration problem instead of incorrectly claiming the file is missing.
  • Invalid settings such as negative patience, non-positive image sizes, invalid worker counts, and non-finite values are rejected earlier.

• 📐 Improved export and dataset behavior

  • Fixed-shape exported models consistently retain their exported image size across repeated predict() calls.
  • TAR dataset extraction now returns the dataset’s actual top-level directory, matching ZIP behavior.
  • Semantic mask caches correctly handle changes to nc=1, preserving foreground pixels in binary mask training.
  • Standalone validation skips labels for classes unsupported by the loaded model instead of crashing.
  • COCO size-specific metrics now remain associated with the correct task, including detection, segmentation, and pose.

• 📦 Smoother dependency installation

  • Corrected Python-version markers and package conflicts for several export extras, improving uv lock and uv sync reliability.

🎯 Purpose & Impact

• Faster deployment: TensorRT users can expect lower inference latency without changing model predictions or application code. ⚡
• Better YOLOE flexibility: YOLOE-26 prompt-free models can now switch between NMS and NMS-free inference as intended.
• Stronger Apple Silicon support: CPU and MPS users benefit from faster validation, semantic segmentation, and training workflows.
• Fewer confusing failures: Configuration, resume, export, dataset, and validation issues are detected earlier with clearer behavior.
• More dependable metrics: COCO evaluation results for boxes, masks, and keypoints are no longer accidentally overwritten or rejected.
• Easier experimentation: Improved dependency resolution and archive handling make setup and dataset workflows more predictable. 🚀

What's Changed

• Disable NNPACK conv2d for CPU inference and training on Apple silicon by @raimbekovm in #26221
• Fix stale semantic mask cache when nc changes to 1 by @cainiao33 in #26227
• Fix check_resume masking real errors as FileNotFoundError by @cainiao33 in #26225
• Keep fixed-shape export imgsz on repeated predict() calls by @cainiao33 in #26241
• Replace box_iou prod reductions with slice products for faster MPS validation by @raimbekovm in #26228
• Fix coco_evaluate writing mask size mAP over the box values and raising on pose by @Nicholas022400701 in #26246
• Fix uv sync and uv lock extra resolution by @hylreg in #26219
• Fix standalone detect val crashing on labels with class index >= model nc by @Nicholas022400701 in #26240
• Speed up TaskAlignedAssigner on CPU and MPS by removing size-2 and middle-axis reductions by @raimbekovm in #26234
• Convert boxes and keypoints once in Results.save_txt, plot and summary by @raimbekovm in #26233
• Fix slow and crashing semantic segmentation on MPS by @Y-T-G in #26239
• Ignore AutoBatch probe failures below the first successful size by @fcakyon in #26224
• Support dual one2many/one2one LRPC heads for prompt-free YOLOE inference by @ShuaiLYU in #26201
• Fix close_mosaic never firing when epochs < close_mosaic by @cainiao33 in #26226
• Reject negative patience at cfg by @wizzseen in #26236
• Floor imgsz at the model stride in check_imgsz by @cainiao33 in #26243
• Fix tar dataset extraction returning the datasets root directory by @cainiao33 in #26244
• Speed up TensorRT FP16 and INT8 engines up to 20% by @Y-T-G in #26223

New Contributors

• @wizzseen made their first contribution in #26236
• @hylreg made their first contribution in #26219

Full Changelog: v8.4.156...v8.4.157

🌟 Summary

v8.4.156 improves remote NDJSON dataset reliability and makes INT8 TensorRT exports faster while preserving accuracy. 🚀

📊 Key Changes

• 🔄 Reliable remote NDJSON conversion — PR #26242 by @cainiao33

  • Refreshes the remote NDJSON manifest inside the existing conversion lock, preventing stale dataset manifests from being silently reused.
  • Uses a private temporary download directory so similarly named manifests from different URLs cannot overwrite one another.
  • Keeps the normalized content hash as the conversion cache:
    • Unchanged datasets—including those with rotated signed URLs—reuse existing labels and images without unnecessary downloads.
    • Changed manifests generate a new dataset directory, preserving files used by active training jobs.
  • Removes NDJSON-specific deletion behavior from the general check_file utility, making file handling more predictable.
  • Validated across detection, classification, depth, local sources, concurrent downloads, interrupted conversions, and URL basename collisions.

• ⚡ Faster and more accurate INT8 TensorRT exports — PR #26171 by @Y-T-G

  • QAT now quantizes more of the model head while keeping accuracy-sensitive final output layers and DFL operations in floating point.
  • BatchNorm layers are fused during export, prediction, and validation, reducing runtime overhead.
  • Quantization ranges are adjusted when BatchNorm is fused so INT8 values remain effectively unchanged.
  • PTQ exports now keep the most accuracy-sensitive head operations out of INT8 and use FP16 where appropriate.
  • TensorRT precision handling was refined for both newer and older TensorRT versions.

• 🧪 More robust testing — PR #26217 by @glenn-jocher

  • A restricted-load model test now uses tolerant numerical comparison rather than requiring bit-for-bit identical outputs across all hardware backends.
  • This reduces false failures without weakening validation of model weights, criteria, or fused layers.

• 🛠️ Maintenance and documentation

  • Updated the self-hosted GitHub Actions cleanup runner from v1.4.39 to v1.4.40.
  • Refreshed the embedded video in the Security Alarm System guide.
  • Updated the export documentation to describe the more precise QAT head behavior.

🎯 Purpose & Impact

• More trustworthy dataset updates: Remote NDJSON publishers can change manifests without users accidentally training on outdated data. ✅
• Less unnecessary network and disk usage: Identical datasets are reused efficiently, even when signed image URLs rotate.
• Safer ongoing training: Changed datasets receive separate directories, so existing jobs continue using the files they started with.
• Improved INT8 deployment performance: TensorRT engines should export and run more efficiently while retaining accuracy-sensitive operations in higher precision. ⚡
• Better cross-platform stability: Testing is less likely to fail because of harmless backend-level numerical differences, particularly on embedded and Jetson environments.
• No major user-facing API changes: The release primarily improves conversion reliability, deployment efficiency, and development infrastructure.

What's Changed

• Speed up QAT and PTQ INT8 TensorRT engines by @Y-T-G in #26171
• Compare restricted-load outputs with allclose in test_restricted_load_criterion by @glenn-jocher in #26217
• Bump eviden-actions/clean-self-hosted-runner from v1.4.39 to v1.4.40 in /.github/workflows by @UltralyticsAssistant in #26229
• Add https://youtu.be/ij9cnYM0EVE to docs by @RizwanMunawar in #26235
• Refresh remote NDJSON at the conversion owner by @cainiao33 in #26242

Full Changelog: v8.4.155...v8.4.156

🌟 Summary

Ultralytics 8.4.155 improves dataset-cache safety, export reliability, training validation, and platform compatibility—helping YOLO26 workflows fail less often and provide clearer feedback. 🚀

📊 Key Changes

• 🧹 Prevented stale labels.cache reuse after dataset configuration changes (PR #26207, the primary release change)
  Dataset cache validation now includes important scan settings such as:

  • Number of classes
  • Keypoint configuration
  • Detection versus pose usage
  • Single-class mode
    This ensures changes to data.yaml trigger a fresh label scan instead of reusing incompatible cached labels.

• ⛔ Invalid epoch values are now rejected early
  Training configurations with epochs=0 or negative values now raise a validation error instead of silently running 100 epochs or producing invalid training outputs.

• 🖥️ Improved Windows OpenVINO inference
  Windows CPU inference now explicitly requests FP32 precision to avoid reduced-precision kernel failures on affected systems.

• 📦 Better control over automatic dependency installation
  Setting YOLO_AUTOINSTALL=False now also prevents automatic apt installations and Edge TPU compiler setup. Missing dependencies produce a warning or actionable error instead of modifying the environment unexpectedly.

• 📱 Fixed YOLO26 pose training on Apple MPS
  The RLE pose-loss weights now use the correct float32 type, allowing YOLO26 pose training to run on MPS devices.

• ⚡ Faster font checking
  Font lookup uses Matplotlib’s cached font list before rescanning the operating system, significantly reducing startup time on systems such as macOS.

• 📤 Improved export and inference paths

  • LiteRT conversion now explicitly uses evaluation mode.
  • MNN examples no longer clamp pixel-coordinate boxes to [0, 1], fixing collapsed bounding boxes and producing correct detections.
  • OBB angle data is passed directly during decoding, avoiding unnecessary model attributes and export warnings.

• 🔧 Updated distributed training compatibility
  DDP now uses the appropriate buffer-synchronization option for newer PyTorch versions, reducing deprecation warnings while preserving behavior.

• ☁️ Updated Platform SDK documentation
  Documentation now recommends ultralytics-platform>=0.1.45 and explains how training metrics, checkpoints, arguments, and host information synchronize with the Ultralytics Platform.

• 🌐 Expanded repository mirroring infrastructure
  A scheduled workflow now mirrors public Ultralytics repositories to GitLab daily, including branches, tags, Git LFS objects, metadata, and repository avatars.

• 📚 Refreshed heatmaps documentation
  The heatmaps guide now features an updated YOLO26 tutorial video.

🎯 Purpose & Impact

• More reliable dataset changes: Users can safely modify class counts, keypoint settings, or task configuration without unknowingly training from incompatible cached labels. Errors are detected during scanning rather than appearing later as confusing loss or indexing failures. 🛡️
• Safer training behavior: Invalid epoch settings now fail clearly at configuration time, preventing accidental long training runs and misleading output files.
• Better device and backend support: YOLO26 pose training on Apple hardware, Windows OpenVINO inference, MNN inference, LiteRT export, and OBB export workflows are more dependable. 💻
• Faster startup on macOS: Cached font discovery can substantially reduce the overhead of validation and training commands that generate plots. ⚡
• Greater environment control: Production, CI, and container users can disable automatic package installation while still receiving useful dependency guidance.
• Improved maintainability: Cleaner export code, fewer PyTorch warnings, and updated Platform integration guidance make the package easier to operate and integrate.

What's Changed

• Add https://youtu.be/UU87gGC-kKg to docs by @RizwanMunawar in #26203
• Add daily GitLab mirrors and fix Windows OpenVINO inference by @glenn-jocher in #26206
• Fix CI warnings at their source: OBB angle attribute, DDP broadcast_buffers, LiteRT training mode by @glenn-jocher in #26215
• Use matplotlib's cached font list in check_font instead of rescanning system fonts by @raimbekovm in #26216
• Fix YOLO26 pose training on MPS by declaring RLE_WEIGHT as float32 by @raimbekovm in #26212
• Document Platform SDK training callbacks and raise the SDK floor to 0.1.45 by @raimbekovm in #26205
• Remove the [0, 1] box clamp from the MNN inference examples by @raimbekovm in #26204
• Reject epochs < 1 at config validation instead of silently training 100 epochs by @cainiao33 in #26208
• Respect YOLO_AUTOINSTALL in apt and Edge TPU install paths by @Y-T-G in #26214
• Fix stale labels.cache reuse after data.yaml semantic edits by @cainiao33 in #26207

Full Changelog: v8.4.154...v8.4.155

🌟 Summary

v8.4.154 improves CoreML export and inference reliability, restores accurate RT-DETR INT8 deployment, reduces training overhead, and strengthens dataset and Platform workflows. 🚀

📊 Key Changes

• 🛠️ CoreML dynamic export fixed — PR #26199

  • YOLO detection, segmentation, pose, and OBB models can now export with dynamic=True without triggering a coremltools arange conversion error.
  • Static CoreML models now correctly process batches of multiple images instead of running inference only on the first image.
  • Supports proper output stacking for raw predictions, embedded NMS, segmentation, and classification models.
  • CoreML export documentation now clarifies restrictions for dynamic inputs, NMS, classification, RT-DETR, and batch sizes.

• ⚡ RT-DETR OpenVINO INT8 export repaired

  • Keeps the RT-DETR decoder in floating point while applying NNCF transformer-aware quantization.
  • Reported RT-DETR-L accuracy improved from approximately 0.0002 to 0.6513 mAP50-95, with nearly unchanged CPU inference speed.

• 🏎️ Faster training, especially on GPUs

  • Avoids unnecessary memory initialization, activation copies, host-device synchronization, and repeated EMA state reconstruction.
  • Enables fused Adam and AdamW optimizers where supported.
  • A measured YOLO26x COCO training step on a B200 GPU improved from 228.9 ms to 183.5 ms—about a 1.25× speedup in that test environment.

• 🧠 More memory-efficient SAM3 mask processing

  • Large semantic masks are upscaled in bounded chunks rather than all at once.
  • This prevents multi-gigabyte temporary allocations while preserving mask results.

• 🎯 Classification validation now prevents class-index mistakes

  • Validation and training splits are aligned to the model’s class names instead of relying on each folder’s local alphabetical ordering.
  • Classes missing from the model are skipped with a warning, preventing silently incorrect accuracy and model-selection metrics.

• 📡 Platform training callbacks now use the Platform SDK

  • Replaces duplicated raw HTTP and retry logic with the generated SDK.
  • Adds controlled POST retries while preserving authentication handling, cancellation, checkpoint signing, payload sanitization, and quiet console-error behavior.
  • Requires ultralytics-platform>=0.1.45.

• ✅ Clearer dataset and validation checks

  • Segment datasets now reject box-only labels or mismatched polygon and box counts.
  • Pose validation reports an actionable error when kpt_shape is missing, including when stale label caches are present.
  • save_json=True now reports small-, medium-, and large-object mAP on detection datasets using faster-coco-eval.
  • Training resume behavior is documented: the checkpoint dataset is restored unless an explicit data= override is provided.

• 🌐 Platform workflow and documentation updates

  • Documents semantic PNG mask imports, similar-image search, generated image variations, model moves between projects, remembered training settings, and verified dataset uploads.
  • Clarifies API rate limits, exact dataset slugs, upload integrity checks, and CoreML limitations.

🎯 Purpose & Impact

• Apple users can export and run models more reliably with dynamic CoreML inputs and multi-image inference now functioning as expected. 🍎
• Deployment accuracy improves for RT-DETR on Intel hardware, making OpenVINO INT8 a more practical option.
• GPU training can be faster and more efficient, particularly for larger workloads, although the reported speedup was measured on a specific B200 setup and may vary by hardware.
• Large-image SAM3 workloads use less peak memory, reducing the risk of out-of-memory errors.
• Classification metrics become trustworthy when train, validation, and test folders contain different class sets.
• Dataset and pose errors are easier to diagnose, reducing time spent tracking down malformed labels or configuration issues.
• Platform training integrations are more maintainable and resilient through centralized SDK behavior and configurable retries. 🔒

What's Changed

• Use Platform SDK for training callbacks by @JaviChulvi in #26190
• Document segment label rejection, resume data override, size mAP with save_json and CoreML dynamic limits by @raimbekovm in #26198
• Upscale SAM3 semantic masks in memory-bounded chunks by @BSchilperoort in #26200
• Fix RT-DETR OpenVINO INT8 export by keeping its decoder float and quantizing with NNCF by @onuralpszr in #26191
• Mirror the trainer's missing-kpt_shape error in standalone pose val by @cainiao33 in #26193
• Document Platform similar image search, mask import, model moves and verified uploads by @raimbekovm in #26195
• [perf] yolo26x COCO training on B200: 228.9 to 183.5 ms per step at 1 GPU, mostly from deterministic mode filling every new tensor by @TarzanZhao in #26180
• Fix silent wrong-class scoring when classification split classes differ from the model by @cainiao33 in #26192
• Fix CoreML dynamic anchor export and static multi-image inference, release 8.4.154 by @raimbekovm in #26199

New Contributors

• @TarzanZhao made their first contribution in #26180
• @BSchilperoort made their first contribution in #26200

Full Changelog: v8.4.153...v8.4.154

🌟 Summary

Ultralytics 8.4.153 delivers a critical SAM3 initialization fix, improves validation reliability, clarifies INT8 export behavior, and expands Ultralytics Platform documentation and annotation capabilities. 🚀

📊 Key Changes

• 🛠️ Fixed SAM3 compile=False handling — PR #26189 by @glenn-jocher

  • Converts the predictor option correctly before passing it to the SAM3 builder:
    • False disables compilation.
    • True enables the default compilation mode.
    • Explicit compilation mode strings are preserved.
  • Prevents unintended torch.compile activation and the resulting PyTorch 2.14 Dynamo error during SAM3 initialization.
  • Validated across all six Platform SAM models for HTTP and WebSocket inference.
  • Package version bumped to 8.4.153.

• ✅ Improved standalone semantic segmentation validation

  • Polygon-based semantic datasets now receive their background class metadata during dataset construction.
  • Prevents background pixels from being incorrectly assigned to the final real class.
  • Makes standalone validation metrics such as mIoU and pixel accuracy meaningful again.

• 🖼️ Fixed standalone classification validation with extra classes

  • Validation now filters samples whose class IDs are outside the model’s class range, matching existing training behavior.
  • Prevents IndexError crashes when the validation dataset contains more classes than the checkpoint supports.
  • Clear warnings identify skipped samples and mismatched classes.

• 🔄 Preserved explicit datasets when resuming training

  • train(resume=True, data="...") now correctly honors the user-provided dataset instead of silently using the checkpoint’s original dataset.
  • Path-based resumes also avoid injecting an unintended task-default dataset.

• ⚙️ Improved INT8 export behavior and documentation

  • Quantization-aware training checkpoints can export their stored INT8 ranges without calibration data for ONNX and TensorRT.
  • RKNN INT8 validation now correctly applies to detection models only, including INT8-only Rockchip targets.
  • RKNN examples now use a compatible detection model and dataset.
  • TensorRT documentation no longer claims that dynamic is automatically enabled for INT8 exports.

• 🎯 Preserved conf=0.0 during IMX export

  • An explicit zero confidence threshold is no longer replaced with 0.001.
  • Added coverage for both conf=0.0 and the default conf=0.25 behavior.

• 🧩 Expanded Platform annotation and billing documentation

  • Added documentation for dataset-wide batch annotation, class mapping, progress tracking, stopping runs, version snapshots, and billing.
  • YOLO smart annotation now includes pose datasets, while SAM remains available for detection, segmentation, semantic, and OBB tasks.
  • Documented the Platform referrals tab, endpoint uptime charges, batch annotation pricing, supported dataset licenses, and Python 3.11 requirement.
  • Agents documentation now reflects that the feature no longer requires Early Access.

• 📚 Documentation and links updated

  • Repository banners now link to the YOLO27 preview and waitlist.
  • Added Huawei Ascend and Google Coral Edge TPU to the Hailo hardware comparison.
  • Removed the unreachable retired Neptune integration page.

🎯 Purpose & Impact

• More reliable SAM3 deployments: Users on newer PyTorch versions can initialize and run SAM3 without unexpected compilation or Dynamo failures. 🧠
• More trustworthy validation results: Semantic and classification validation now better reflect the actual model and dataset configuration.
• Safer resume workflows: Explicit dataset overrides are respected, reducing the risk of accidentally continuing training on the wrong data.
• Simpler edge deployment: INT8 export rules are clearer and better aligned with actual ONNX, TensorRT, and RKNN behavior.
• More predictable export thresholds: IMX users can intentionally disable filtering with conf=0.0.
• Broader Platform workflows: Pose smart annotation and batch annotation make dataset preparation faster, while updated billing and integration documentation makes Platform usage easier to understand.
• Upgrade recommendation: Users relying on SAM3, standalone validation, training resume, IMX export, or RKNN INT8 export should upgrade to 8.4.153. 🔄

What's Changed

• Update banner destination links to YOLO27 by @raimbekovm in #26177
• Align INT8 export, Platform Python floor and Hailo comparison docs with merged behavior by @raimbekovm in #26174
• Remove the unreachable Neptune integration page by @raimbekovm in #26170
• Fix standalone semantic val mislabeling background pixels as the last class by @cainiao33 in #26176
• Preserve an explicit data= override when resuming training by @cainiao33 in #26175
• Link the YOLO27 waitlist from the coming-soon notice by @raimbekovm in #26188
• Align Platform docs with dataset licenses, the referrals tab and pose smart annotation by @raimbekovm in #26185
• Document Platform batch annotation and its billing by @raimbekovm in #26186
• Fix the RKNN INT8 detect-only gate for INT8-only targets and its docs example by @raimbekovm in #26184
• Preserve conf=0.0 in IMX export score_threshold by @cainiao33 in #26183
• Fix standalone classify val crash on datasets with extra classes by @cainiao33 in #26182
• Fix SAM3 compile=False handling and release 8.4.153 by @glenn-jocher in #26189

Full Changelog: v8.4.152...v8.4.153

🌟 Summary

v8.4.152 improves system monitoring by adding cached NVIDIA driver and driver-supported CUDA version information to SystemLogger.

📊 Key Changes

• 🖥️ Added top-level driver_version and cuda_version fields to system metrics collected on NVIDIA-enabled systems.
• ⚡ Version information is queried once when SystemLogger starts, avoiding repeated NVML calls during metric collection.
• 🛡️ GPU utilization monitoring continues to work even when version information is unavailable; missing fields are simply omitted.
• 📌 cuda_version represents the CUDA version supported by the installed NVIDIA driver, matching the value reported by nvidia-smi—not the installed CUDA toolkit or PyTorch build version.
• 📦 Bumped the package version from 8.4.151 to 8.4.152.

🎯 Purpose & Impact

• 🔍 Monitoring systems can now receive NVIDIA driver and CUDA compatibility details directly from SystemLogger, without collecting them separately.
• 🚀 Caching keeps repeated metric collection efficient, with validation confirming only one query per version per logger instance.
• 🌍 Non-NVIDIA and macOS environments remain compatible, with NVIDIA-specific fields omitted when unavailable.
• ✅ Testing on an eight-GPU server confirmed accurate versions, reliable collection across cumulative and rate modes, and strong performance.

What's Changed

• Report NVIDIA driver and CUDA versions in SystemLogger by @glenn-jocher in #26173

Full Changelog: v8.4.151...v8.4.152

🌟 Summary

Ultralytics 8.4.151 improves export correctness, TensorRT QAT performance, tracking behavior, model diagnostics, and documentation—while introducing no new runtime arguments. 🚀

📊 Key Changes

• Preserved conf=0.0 in NMS exports 🎯

  • The most important fix ensures an explicit zero confidence threshold is written into exported ONNX and other NMS-enabled graphs.
  • The default confidence of 0.25 is now applied only when conf is omitted or set to None.
  • Added an ONNX regression test that verifies the serialized threshold.

• Faster QAT TensorRT engines ⚡

  • Quantization-aware training exports now run unquantized TensorRT layers in FP16 instead of FP32.
  • Applies to TensorRT 10 and TensorRT 11 while preserving INT8 quantization ranges.
  • Reported benchmarks show QAT INT8 performance improving substantially, with no changes to standard post-training INT8 or FP16 exports.

• Explicit zero confidence preserved in tracking 🎥

  • model.track(conf=0.0) now keeps the requested value instead of replacing it with the default 0.1.
  • Omitted confidence values still default to 0.1, maintaining existing tracking behavior.

• Warnings for blank class labels ⚠️

  • AutoBackend now warns when model class names are empty or contain only whitespace.
  • The original names remain unchanged, and repeated access does not produce repeated warnings.

• Expanded tracking and validation documentation 📚

  • Solution guides now document all forwarded tracking arguments, including imgsz, max_det, and quantize.
  • Validation documentation now explains the fraction option for selecting subsets of training, validation, or test data.
  • Corrected the documented YOLO26n-depth parameter count.

• Improved PyTorch hook guidance 🧩

  • Added a runnable example showing how to register forward hooks after the training model is initialized.
  • Clarifies behavior with distributed training, EMA checkpoints, validation, prediction, and tracking.

• Updated YOLO27 preview documentation 🔭

  • Reorganized the preview page, clarified NMS and NMS-free detection heads, removed unsupported performance claims, and linked the preview from relevant YOLO26, task, and mode pages.
  • YOLO27 remains unreleased; YOLO26 continues to be the recommended stable model family.

• Documentation rendering and layout fixes 🖼️

  • Corrected isolated-object images that previously displayed raw Markdown syntax.
  • Shortened an Intel DL Streamer heading to prevent table-of-contents overflow.

🎯 Purpose & Impact

• More predictable exports: Users can intentionally export with conf=0.0 without the value being silently changed to 0.25. ✅
• Better deployment speed: QAT TensorRT engines should make more effective use of FP16 hardware acceleration, particularly on GPUs where FP32 fallback was slowing inference.
• More reliable tracking controls: Confidence thresholds now behave consistently when users explicitly provide zero or other numeric values.
• Easier troubleshooting: Warnings for blank class names help identify why detections may appear without visible labels.
• Clearer documentation: Expanded argument tables, hook examples, validation guidance, and YOLO27 status notes make the platform easier to use and reduce ambiguity for both new and experienced developers. 📖

What's Changed

• Clarify YOLO27 preview documentation and lead with the page heading by @glenn-jocher in #26162
• Render the isolated-object figures as HTML and shorten the DL Streamer heading by @raimbekovm in #26168
• Document the forwarded solution track args, val fraction and the fused depth params by @raimbekovm in #26166
• Link the YOLO27 preview from the task, mode and YOLO26 pages by @raimbekovm in #26165
• Warn on empty class names when initializing AutoBackend by @cainiao33 in #26164
• Preserve explicit zero confidence in track mode by @aswanth-07 in #26161
• Document forward hooks across training model rebuilds by @aswanth-07 in #26157
• Run the float head of QAT TensorRT engines in FP16 by @Y-T-G in #26167
• Preserve conf=0.0 in nms=True exports instead of baking 0.25 into the graph by @cainiao33 in #26163

Full Changelog: v8.4.150...v8.4.151

🌟 Summary

v8.4.150 improves secure checkpoint loading, restores compatibility with fused YOLOE models, and adds performance and documentation updates across Ultralytics. 🚀

📊 Key Changes

• ⚡ Faster restricted checkpoint loading — PR #26155 by @glenn-jocher

  • Resolves only the model classes actually referenced by a checkpoint instead of importing and scanning every supported model module.
  • Preserves existing trusted namespaces, aliases, registration locks, and allow-list behavior.
  • Restores the safe forward and forward_fuse bindings required by fused YOLOE checkpoints.
  • Improves error reporting by showing the rejected object type without calling repr() on partially reconstructed modules.
  • Adds support for legacy NumPy metadata and embedded Ultralytics CLIP objects during restricted loading.
  • Extends tests to cover legacy checkpoints, fused models, NumPy fitness values, and inference-output consistency.

• 🛡️ Safer diagnostics and Ray Tune compatibility — PR #26151

  • Prevents Sentry from capturing command-line arguments and local variables that may contain credentials.
  • Enforces Ray version 2.41.0 or newer and removes obsolete compatibility code for older Ray releases.

• 📈 More efficient RT-DETR training on dense datasets — PR #26150

  • Avoids calculating matching costs between predictions and targets from different images.
  • Reduces unnecessary cost storage and padding overhead, with the greatest benefits expected on datasets containing many objects per image, such as SKU-110K.
  • Keeps the assignment solver, losses, precision, and training configuration unchanged.

• 📚 YOLO27 preview documentation

  • Publishes a clearly labeled YOLO27 Coming Soon model page with planned architecture details, supported tasks, preliminary benchmarks, and future usage examples.
  • Adds discovery links throughout the README and documentation.
  • Clarifies that YOLO27 weights, package support, and a launch date are not currently available; YOLO26 remains the recommended released model.

• ☁️ Expanded Ultralytics Platform deployment documentation

  • Documents configurable CPU and memory resources, uptime billing, scale-to-zero behavior, and lightweight monitoring for eligible paid endpoints.
  • Explains that prediction examples and statistics are temporary and may be lost after stopping, restarting, redeploying, resizing, or replacing an endpoint.
  • Describes saving useful prediction examples to datasets for persistent storage.

• 🏆 Clearer model benchmark tables

  • Standardizes bold highlighting for best accuracy, speed, latency, size, parameter count, and computational-cost values across model documentation.
  • Updates benchmark tables for YOLO26, YOLO11, YOLOv8, YOLOv9, YOLO12, YOLOE, SAM, YOLO-NAS, and other supported families without changing the underlying results.

🎯 Purpose & Impact

• More reliable production loading: Fused YOLOE checkpoints and older Platform checkpoints can load and warm up successfully again. ✅
• Faster cold starts: Restricted loading performs less unnecessary module discovery, which should reduce startup overhead, especially in deployment environments.
• Maintained security: Loading remains restricted to trusted, explicitly permitted classes and bindings; the changes do not broadly enable arbitrary checkpoint contents.
• Better dense-object training performance: RT-DETR users working with crowded scenes may see faster training and lower memory usage, while ordinary datasets should see little or no change.
• Improved privacy: Sentry error reports are less likely to contain credentials supplied through yolo login or other local variables.
• Clearer upgrade guidance: Users can explore planned YOLO27 capabilities while understanding that it is not yet available. For current projects, use YOLO26 or deploy through the Ultralytics Platform.

What's Changed

• Docs: Deploy configuration and lightweight monitoring by @sergiuwaxmann in #26149
• Prevent Sentry credential capture and remove obsolete Ray fallback by @glenn-jocher in #26151
• Speed up RT-DETR matching on dense datasets by @fcakyon in #26150
• Support legacy NumPy metadata and CLIP in restricted loading by @glenn-jocher in #26153
• Publish YOLO27 coming-soon documentation and discovery links by @glenn-jocher in #26154
• Highlight best benchmark values across model documentation by @glenn-jocher in #26156
• Speed up restricted loading and restore fused checkpoints by @glenn-jocher in #26155

Full Changelog: v8.4.149...v8.4.150

🌟 Summary

Ultralytics 8.4.149 improves training reliability, export compatibility, and Platform documentation—especially for resumed runs and third-party integrations. 🚀

📊 Key Changes

• More reliable training resume behavior 🔄

  • Resuming from a checkpoint now preserves the dataset stored in that checkpoint instead of unexpectedly switching to the default COCO8 dataset.
  • An explicitly provided data= argument still takes priority.
  • Missing datasets now produce the expected dataset-not-found error rather than causing optimizer or model-shape failures.
  • Resuming checkpoints created with AMP disabled no longer attempts to load an empty AMP scaler state.

• Safer ONNX CUDA fallback ⚙️

  • CUDA availability is now checked after ONNX Runtime creates its inference session.
  • If the CUDA execution provider is unavailable, inference automatically falls back to CPU with a clear warning instead of crashing.

• Ray Tune compatibility improvements 📈

  • Ray Tune integration now verifies that a compatible Ray version is installed.
  • Training metrics are reported through the current Ray Tune API, reducing failures with newer Ray releases.

• More robust Weights & Biases runs 📊

  • Automatically generated W&B run names are limited to 64 characters, preventing crashes caused by overly long names.

• Improved invalid-argument errors 🛠️

  • CLI help is logged separately from the syntax error message, helping the core error remain visible even when logs are truncated.

• RT-DETR LiteRT export and inference fixes 📱

  • RT-DETR outputs are no longer incorrectly coordinate-normalized during LiteRT export or inference.
  • This brings LiteRT detection metrics much closer to ONNX results, improving accuracy across Python and JavaScript/npm workflows.

• SAM 3.1 becomes the default Platform smart annotation model ✨

  • Ultralytics Platform now lists SAM 3.1 as the default smart annotation model, while SAM 3 and SAM 2.1 remain available.
  • SAM 3.1 image prediction support and current limitations, including unsupported Object Multiplex video tracking, are documented.

• Platform AutoTrain documentation reorganized 📚

  • AutoTrain has moved to Platform → Train → AutoTrain.
  • The documentation now emphasizes baseline creation, controlled comparisons, experiment review, billing, and team workspace usage.

• Package version updated 📦

  • Version bumped to 8.4.149.

🎯 Purpose & Impact

• Fewer training crashes: Users can resume interrupted jobs more confidently, even when datasets have moved or AMP settings differ.
• Better deployment resilience: ONNX models requested on CUDA systems can continue running on CPU when CUDA support is unavailable.
• More dependable integrations: Ray Tune and W&B workflows are less likely to fail because of API differences or naming limits.
• Improved edge deployment accuracy: RT-DETR models exported to LiteRT should produce results more consistent with ONNX, benefiting mobile and browser applications.
• Simpler annotation workflows: New users on Ultralytics Platform get SAM 3.1 by default for high-quality image annotation, with alternative SAM models available for speed or compatibility.
• Clearer experimentation guidance: The reorganized AutoTrain documentation makes it easier to run repeatable training experiments and choose the next improvement step.

What's Changed

• List SAM 3.1 as the default Platform smart annotation model by @glenn-jocher in #26145
• Move AutoTrain documentation under Platform Train by @JaviChulvi in #26146
• Fix normalization conditions in DETR models for LiteRT export and inference by @onuralpszr in #26142
• Fix resume, ONNX CUDA fallback, Ray Tune and W&B training crashes by @glenn-jocher in #26148

Full Changelog: v8.4.148...v8.4.149

🌟 Summary

Ultralytics 8.4.148 makes SAM 3.1 image prediction checkpoints load correctly, enabling reliable point, box, text, and exemplar-based segmentation. 🎯

📊 Key Changes

• SAM 3.1 checkpoint support 🧠

  • Added checkpoint-key mapping so sam3.1_multiplex.pt works correctly with SAM and SAM3Predictor.
  • Fixes an issue where SAM 3.1’s prompt encoder, mask decoder, embedding, and neck could be silently initialized with random weights.
  • Existing sam3.pt checkpoints remain compatible; the key mapping is effectively a no-op for SAM 3.

• Supported SAM 3.1 prediction workflows 🖼️

  • Point and box prompts through SAM and SAM3Predictor.
  • Text and image-exemplar prompts through SAM3SemanticPredictor.
  • SAM 3.1 video tracking with the new Object Multiplex memory architecture is not yet ported. Use the original sam3.pt with the SAM 3 video predictors for video workflows.

• Documentation updates 📚

  • Added SAM 3.1 guidance, supported scope, benchmark comparisons, and gated-weight download instructions.
  • Updated the model index to list sam3.1_multiplex.pt.

• Repository guidance improvements 🛠️

  • Streamlined the root AGENTS.md and moved documentation-specific instructions into a new docs/AGENTS.md.
  • Preserved important development, testing, release, and implementation-boundary guidance while reducing duplicated repository walkthroughs.

• Version update 📦

  • Bumped the package version from 8.4.147 to 8.4.148.

🎯 Purpose & Impact

• ✅ Correct segmentation results: SAM 3.1 image predictors now use the trained checkpoint components instead of silently falling back to randomly initialized prompt-processing layers.
• 🚀 Easier SAM 3.1 adoption: Users can load the new checkpoint through the existing Ultralytics SAM interfaces with minimal workflow changes.
• 🔍 Clearer expectations: Documentation distinguishes supported SAM 3.1 image prediction from unsupported Object Multiplex video tracking.
• 🔐 Access requirement: SAM 3.1 weights remain gated by Meta and must be obtained from the approved SAM 3.1 model page on Hugging Face.
• 👩‍💻 Improved contributor experience: More focused agent guidance makes repository development instructions easier to discover without changing end-user model behavior.

What's Changed

• Trim derivable content from AGENTS.md and move docs guidance to docs/AGENTS.md by @glenn-jocher in #26138
• Focus repository guidance and retain scoped agent instructions by @glenn-jocher in #26141
• ultralytics 8.4.148 Load SAM 3.1 checkpoints into the SAM 3 image predictors by @glenn-jocher in #26144

Full Changelog: v8.4.147...v8.4.148

🌟 Summary

Ultralytics v8.4.147 improves preprocessing for grayscale images, strengthens SAM segmentation behavior, and refreshes Platform, installation, and CI documentation. 🚀

📊 Key Changes

• 🖼️ Albumentations now supports grayscale images

  • Albumentations preprocessing runs on images with either 1 channel (grayscale) or 3 channels (color).
  • Images with unsupported channel counts continue to be skipped.
  • This is the primary change in PR #26133 by @onuralpszr.

• 🎯 SAM generation improvements

  • Added min_mask_region_area to SAM generation to remove small mask regions and holes.
  • Corrected crop prompt scaling so prompts cover the full resized crop rather than only part of it.
  • Improved confidence filtering so direct generate() calls behave consistently with normal prediction.
  • Standardized source image shape handling for both image arrays and tensors.
  • Updated SAM documentation with an example using min_mask_region_area.

• 📚 Expanded Platform and agent documentation

  • Added documentation for Platform Agents, including visual workflows, conditions, dataset collection, Slack alerts, execution options, and run monitoring.
  • Added Ask AI Autotraining guidance for dataset exploration, training, model comparison, annotation, export, and deployment.
  • Documented the new platform-cli agent skill and ul cloud workflows.
  • Added guidance for configurable deployment resources and current API-key behavior.

• ⚡ Improved installation and quickstart guidance

  • Reworked the quickstart around installing Ultralytics and running a first prediction with pretrained YOLO26.
  • Moved persistent settings into a dedicated Settings page.
  • Added clearer examples for CLI, Python, Docker, Conda, headless environments, and development installations.

• 🛡️ More reliable downloads

  • safe_download() now makes curl fail on HTTP errors instead of saving an error page as if it were a valid file.
  • This helps surface failed downloads immediately.

• 🧪 Stronger CI and runner environments

  • Re-enabled trusted GPU CI using CUDA 13.2 and TensorRT CUDA 13.
  • Added retry handling for transient APT installation failures.
  • Added GitHub CLI (gh) and jq to CPU and GPU runner images.

🎯 Purpose & Impact

• Better grayscale workflows: Users working with medical, industrial, document, or scientific grayscale imagery can now apply their configured Albumentations transforms during preprocessing. 🩶
• Cleaner SAM results: Small artifacts can be filtered, crop-based prompting is more accurate, and direct SAM generation results are more consistent with standard prediction behavior.
• More dependable tooling: Failed downloads are reported correctly, while updated CI and runner dependencies reduce avoidable workflow failures.
• Easier onboarding: New users can reach a first YOLO26 prediction faster with a simpler quickstart and clearer installation options.
• More accessible Platform automation: The new documentation makes it easier to use Platform Agents, Ask AI, and the Platform CLI for training, dataset management, export, and deployment workflows. 🤖
• No new model family or weights were introduced in this tag: the release primarily delivers preprocessing fixes, SAM enhancements, reliability improvements, and documentation updates.

What's Changed

• Retry apt installs in the isolated export job by @raimbekovm in #26122
• Add https://youtu.be/ZsxHIN2PoiY to docs by @RizwanMunawar in #26124
• Fix curl downloads saving HTTP error pages as valid files by @lakshanthad in #26126
• Enable GPU CI on CUDA 13 runners by @glenn-jocher in #26127
• Install GitHub CLI in CPU and GPU runner images by @glenn-jocher in #26131
• Install jq in CPU and GPU runner images by @glenn-jocher in #26134
• Document the Platform CLI agent skill by @JaviChulvi in #26136
• Document Platform Agents visual workflows by @glenn-jocher in #26135
• Document Platform Autotraining with Ask AI by @JaviChulvi in #26137
• Add min_mask_region_area to SAM generate() and scale crop point prompts to the resized crop by @raimbekovm in #26129
• Rewrite quickstart around installation and a first prediction by @raimbekovm in #26128
• Apply Albumentations to grayscale images by @onuralpszr in #26133

Full Changelog: v8.4.146...v8.4.147

🌟 Summary

Ultralytics v8.4.146 improves RT-DETR reliability across small inputs, dynamic exports, and tracking, while updating export portability, Windows training behavior, data downloads, and GPU container environments. 🚀

📊 Key Changes

• 🛠️ RT-DETR inference fixes

  • Prevents failures when small images produce fewer anchors than the configured query count.
  • Keeps query counts, denoising masks, and learned query embeddings consistent.
  • Preserves the correct detection limit in dynamic ONNX and OpenVINO exports.
  • Makes RT-DETR tracking compatible with TrackTrack by avoiding an unsupported NMS recovery argument.
  • Validated across image sizes from 32 to 640 pixels, TorchScript, CoreML, ONNX, OpenVINO, training, validation, and tracking.

• 📦 More device-agnostic TorchScript exports

  • Removes hard-coded device assumptions from export-time tensor creation.
  • TorchScript models exported on CUDA or Apple MPS can now be loaded and used on other devices, such as CPU.
  • Fixes dynamic RT-DETR export behavior so bounding boxes remain correct at sizes different from the export size.

• 🎯 Improved FP16 embedded-NMS inference

  • Converts end-to-end NMS output rows to float32.
  • Prevents segmentation, pose, and OBB FP16 exports from failing during mask or box processing.

• 🪟 Better Windows CUDA training defaults

  • Stops automatically enabling channels_last memory format on Windows, where it could significantly reduce performance.
  • Users can still explicitly enable it with channels_last=True.

• 🌐 More reliable NDJSON image downloads

  • Replaces the previous 30-second total timeout with connection and socket-inactivity timeouts.
  • Large images on slow or shared connections are less likely to fail while downloading.

• 🐳 Updated GPU and export environments

  • AMD64 GPU images now use PyTorch 2.14 with CUDA 13.2.
  • NVIDIA ARM64 images move to NVIDIA PyTorch 26.08 with CUDA 13.4 and updated ONNX Runtime GPU support.
  • GPU runner images now include g++ for compiled training workflows and TensorRT for CUDA 13.
  • LiteRT remains in the Python export image because its current dependencies are not compatible with PyTorch 2.14.
  • Conda Docker builds now use conda-forge PyTorch 2.13, torchvision 0.28, and CUDA 13.0.

• 🧪 CI and maintenance updates

  • Restores the GPU CI workflow while keeping it temporarily disabled through an explicit condition.
  • Adds disk cleanup before isolated export setup.
  • Updates documentation for CUDA 13, JetPack, TensorRT, DLA, Conda images, and export compatibility.
  • Bumps the package version from 8.4.145 to 8.4.146.

🎯 Purpose & Impact

• Greater RT-DETR stability: Small images and dynamically sized inputs should no longer trigger query or anchor-related inference failures. ✅
• More dependable deployment: Models exported on one accelerator can be used on another device more reliably, reducing device-specific TorchScript issues.
• Safer FP16 workflows: Embedded-NMS exports for segmentation, pose, and OBB models now handle post-processing without dtype errors.
• Faster Windows training in common cases: Windows CUDA users avoid an automatic memory layout optimization that was measured to slow training substantially.
• Improved dataset conversion reliability: Large NDJSON image downloads can complete on constrained networks instead of repeatedly stopping at a total timeout.
• Clearer GPU compatibility requirements: CUDA 13 containers may not support older Maxwell, Pascal, or Volta GPUs, and Jetson users must match TensorRT versions to their JetPack release.
• Deployment caution: The updated GPU images require compatible NVIDIA host drivers, and TensorRT engines should be rebuilt or validated when the GPU, CUDA, TensorRT, or JetPack environment changes.

What's Changed

• Update GPU Docker images to current PyTorch and CUDA releases by @glenn-jocher in #26115
• Restore GPU CI with a temporary disable condition by @glenn-jocher in #26119
• Time NDJSON image downloads by socket inactivity instead of a 30 s total deadline by @raimbekovm in #26113
• Keep TorchScript exports device-agnostic by @raimbekovm in #26118
• Cast end-to-end NMS rows to float32 so FP16 exports with embedded NMS decode segment masks by @raimbekovm in #26114
• Skip channels_last auto-enable on Windows CUDA training by @Y-T-G in #26105
• Fix RT-DETR inference for small and dynamic inputs by @glenn-jocher in #26120

Full Changelog: v8.4.145...v8.4.146

🌟 Summary

Improves Platform login reliability in Ultralytics 8.4.145 by reporting authentication and connection failures with a proper non-zero exit status, making command-line automation more dependable. ✅

📊 Key Changes

• 🔐 Login failures now exit with status 1
  • Invalid API keys and rejected authentication requests now terminate with an error instead of only displaying a warning.
  • Network and connection failures also return a failure status.
• 🧭 Consistent behavior across login commands
  • The fix applies to both yolo login and delegated ul login flows.
• ✅ Successful operations are unchanged
  • Successful login continues to exit normally and save credentials.
  • Logout continues to clear credentials and return status 0.
• 📦 Version updated
  • Package version increases from 8.4.144 to 8.4.145 so the login fix is available through PyPI.

🎯 Purpose & Impact

• 🤖 Better automation: Scripts and CI pipelines can now reliably detect failed logins instead of mistakenly treating them as successful.
• 🛠️ Clearer error handling: Authentication and connectivity problems are surfaced as actual command failures while retaining the existing error messages.
• 🔄 Safer credential behavior: Failed login attempts do not overwrite valid saved credentials.
• 🚀 Release availability: Users installing ultralytics==8.4.145 receive the corrected login behavior.

What's Changed

• Return failure status from Platform login by @glenn-jocher in #26116
• Bump to 8.4.145 for login failure exit status by @glenn-jocher in #26117

Full Changelog: v8.4.144...v8.4.145

🌟 Summary

v8.4.144 improves model reliability, numerical stability, inference precision, and deployment workflows without changing model architectures. 🚀

📊 Key Changes

• More reliable model YAML loading 🧩
  Explicitly requested files such as custom26n.yaml are now loaded before attempting a scale-unified fallback like custom26.yaml. This prevents custom model definitions from being silently replaced.

• Safer training and assignment edge cases 🛡️
  Empty-label batches now return correctly typed boolean foreground masks, keeping task-aligned assigners consistent with their normal output contract.

• Stable CIoU calculations for reduced precision 🔢
  The overlap calculation now avoids NaN values and invalid gradients for identical FP16 and BF16 boxes, improving training stability on lower-precision hardware.

• More efficient repeated model.to() calls ⚡
  Cached predictors are preserved when a model is already on the requested device or precision. This avoids unnecessary rebuilding, reducing latency and memory use during repeated inference setup.

• Improved Triton FP32 behavior 🎯
  Triton client tensors now remain in FP32 even when quantize=16 is requested. This prevents unintended input rounding and output conversion on the client while leaving the server-side model precision unchanged.

• Faster semantic segmentation mosaic loading 🖼️
  Semantic masks for buffered mosaic images are now retained in RAM instead of being decoded repeatedly. Benchmarks showed approximately 1.26× to 1.77× faster data loading, depending on the configuration.

• More flexible CLI configuration 🛠️
  Arguments provided before cfg=<file> are now preserved, and blank data values in copied configuration files correctly fall back to the task’s default dataset.

• Broader OpenVINO/NNCF compatibility 📦
  The general upper version limit for NNCF has been removed, allowing newer NNCF 3.x releases with modern PyTorch and OpenVINO installations while retaining legacy restrictions where required.

• More informative quantization documentation 📚
  QAT versus post-training quantization results are now documented for YOLO26 models. The examples show that QAT provides little benefit for the smallest model but can recover substantially more INT8 accuracy for larger models.

• More robust CI and test infrastructure ✅
  CLA permissions were corrected, DEEPX export environments are skipped on machines with less than 15 GiB of RAM, and checkpoint corruption tests now work correctly with channels-last tensors.

• Documentation maintenance 🔗
  The DL Streamer system requirements link was fixed, and the documentation license banner now uses a CDN-hosted asset.

🎯 Purpose & Impact

• Users get more predictable model behavior, especially when working with custom YAML files or unusual training batches.
• Training with FP16 or BF16 is more robust, reducing the risk of silent NaN values that can interrupt training or corrupt gradients.
• Inference pipelines can be faster and lighter, particularly when repeatedly moving models between devices or processing semantic segmentation datasets with mosaic augmentation.
• Triton deployments preserve the intended FP32 data path, helping maintain accuracy when the remote server uses FP32 inference.
• INT8 deployment decisions are easier to make: QAT is most valuable for larger YOLO26 models or cases where calibration causes a noticeable accuracy drop, while smaller models may benefit little from the additional training cost.
• Export and CI workflows are more dependable across hardware and software environments, with clearer handling of memory limitations and newer OpenVINO dependencies.

What's Changed

• Fix broken DL Streamer system requirements link by @onuralpszr in #26096
• Quote the measured INT8 accuracy of QAT against calibration by @Bovey0809 in #26093
• Serve the buffered mosaic tile's semantic mask from ram instead of re-decoding it by @raimbekovm in #26099
• Keep Triton client tensors in FP32 when quantize=16 is requested by @raimbekovm in #26097
• Honor CLI arguments placed before cfg= and a blank data in the cfg file by @raimbekovm in #26092
• Keep the cached predictor when .to() does not move the model by @onuralpszr in #26098
• Write EMA test poison by index so it lands on channels_last CUDA weights by @onuralpszr in #26100
• Remove general NNCF upper version bound by @glenn-jocher in #26104
• Update license banner image to CDN asset by @raimbekovm in #26103
• Fix model loading, numerical edge cases, and CI setup by @glenn-jocher in #26106

Full Changelog: v8.4.143...v8.4.144

🌟 Summary

v8.4.143 introduces INT8 quantization-aware training for YOLO26, improves deployment and evaluation workflows, and delivers a broad documentation and integration refresh. 🚀

📊 Key Changes

• 🧠 INT8 quantization-aware training (QAT) — PR #26083

  • Adds support for quantize=8 during train mode.
  • Fine-tunes pretrained weights while simulating INT8 quantization, allowing the model to adapt before deployment rather than being quantized only after training.
  • Stores calibrated quantization ranges in the checkpoint and supports direct export to ONNX and TensorRT with Q/DQ nodes—without requiring calibration data during export.
  • Reuses the existing quantize argument and adds no new public API or dependency.
  • Particularly useful for deploying YOLO26 models to resource-constrained edge hardware.

• 📈 More complete validation metrics — PR #24489

  • Adds per-class AP75 values to validation summaries, DataFrame outputs, CSV files, and JSON exports.
  • Applies consistently to detection, segmentation, pose, and OBB tasks.

• 🚀 Improved YOLO26 deployment workflows

  • DeepStream documentation now uses the native Ultralytics exporter instead of a separate third-party script.
  • Clarifies when to use the NMS-free head with nms=False and updates Triton, DALI, SAM, and C++ examples accordingly.
  • Adds documentation for Apple’s Core AI export target and backend.
  • Corrects quantization guidance for formats that support only specific precisions.

• 🛠️ Training and model lifecycle fixes

  • Corrects pretrained-model handling after reset_weights(), during resume operations, and in multi-dataset or distributed training.
  • Preserves Platform model URIs when training through Python.
  • Keeps prediction output directories when copying or converting Results objects.
  • Removes stale SAM3 decoder coordinate caches, improving reliability across device and precision changes.

• 🌍 Platform and authentication updates

  • Uses the Ultralytics Platform SDK for login, model downloads, dataset exports, and training resource access.
  • Aligns Python 3.11+ installations with the shared-login SDK requirements while preserving local workflows on older Python versions.

• 📚 Extensive documentation improvements

  • Reviews and corrects examples, CLI syntax, export details, tracker behavior, dataset layouts, image decoding, hardware integrations, and model descriptions.
  • Expands YOLO26 depth-estimation dataset documentation with new FAQs and dataset guidance.
  • Updates Ultralytics Platform links across dataset and training pages.
  • Documents improved Albumentations support for boxes, polygons, keypoints, depth maps, and semantic masks.
  • Refreshes DeepStream, OpenVINO, Core AI, Edge TPU, Ray Tune, Jetson, and other integration guides.

🎯 Purpose & Impact

• ⚡ Better edge performance: QAT can help INT8-deployed models retain more accuracy while benefiting from lower memory usage, faster inference, and reduced power consumption.
• 📦 Simpler deployment: A QAT-trained checkpoint can be exported to ONNX or TensorRT without a separate calibration dataset, reducing deployment steps.
• 📊 Better model analysis: Per-class AP75 makes it easier to identify which categories perform well or need additional training.
• 🔌 Broader hardware support: Updated exporters and documentation make YOLO26 easier to deploy with DeepStream, TensorRT, Core AI, OpenVINO, Platform, and other environments.
• 🧰 More reliable training: Pretrained-state and resume fixes reduce the risk of unintentionally training from random initialization or using the wrong checkpoint.
• ✅ Clearer user experience: Corrected documentation and examples should reduce setup errors, especially for new users and deployment engineers.

What's Changed

• Include per-class AP75 in validation summaries and exports by @woreom in #24489
• Review docs pages for correctness, consistency, and Platform links by @glenn-jocher in #26070
• Drop the FP32 promise from the export quantize hint on formats that cannot export FP32 by @raimbekovm in #26080
• Export YOLO26 for DeepStream with the Ultralytics exporter by @raimbekovm in #26077
• Honor the pretrained argument after reset_weights() and in MultiTrainer by @raimbekovm in #26084
• Preserve prediction save directory in Results copies by @aswanth-07 in #26074
• Select the NMS-free head explicitly where docs and examples still assume it by @raimbekovm in #26072
• Remove stale SAM3 decoder coordinate caches by @tritsystem in #26075
• Update docs for the max_det dataset raise, Core AI export and spatial Albumentations by @raimbekovm in #26078
• Use Platform SDK for login and training resource downloads by @glenn-jocher in #26087
• Require Platform SDK >=0.1.32 for shared login by @glenn-jocher in #26088
• Add INT8 quantization-aware training via quantize=8 in train mode by @Bovey0809 in #26083

New Contributors

• @tritsystem made their first contribution in #26075
• @woreom made their first contribution in #24489

Full Changelog: v8.4.142...v8.4.143

🌟 Summary

Ultralytics v8.4.142 unifies YOLO inference and export behavior under the nms option, making it easier to choose between maximum accuracy and NMS-free speed while improving benchmark consistency and deployment documentation.

📊 Key Changes

• 🔄 Unified end2end into nms

  • nms=None (default) uses the one-to-many head with Ultralytics-managed NMS.
  • nms=True uses the one-to-many head and embeds NMS in supported exports.
  • nms=False selects the one-to-one NMS-free head when available.
  • The legacy end2end argument is deprecated:
    • end2end=True maps to nms=False.
    • end2end=False maps to nms=None, unless explicitly combined with nms=True.
  • Both YOLO26 heads remain trained, while validation, checkpoint selection, and early stopping now follow the selected inference head.

• 📦 Improved model fusion and export handling

  • Model fusion now removes the unused detection branch regardless of which inference head is selected.
  • Exported models preserve their native output behavior, while unsupported formats automatically fall back to compatible raw outputs.
  • NMS options are now documented consistently across ONNX, TensorRT, CoreML, OpenVINO, MNN, Ascend, Hailo, and other integrations.

• 📏 More reliable cross-format benchmarks

  • Benchmark validation now uses square inputs consistently across native and exported models.
  • This ensures accuracy and latency comparisons are based on the same preprocessing rather than mixing rectangular and square inputs.

• 🐛 Training resume fix

  • Resuming training with Albumentations transforms no longer fails when saving arguments to YAML.
  • The fix addresses argument normalization order without changing the training workflow.

• 🌐 Expanded deployment documentation

  • Added a new guide for deploying Ultralytics models on Luxonis OAK RVC2 and RVC4 cameras.
  • Improved Ambarella CVflow deployment guidance, including SDK compilation, host validation, and Cavalry device deployment.
  • Documented reliable DEEPX SDK installation using the vendor wheel repository.

• 🧪 Broader test coverage and maintenance

  • Updated prediction, validation, export, Hailo, and model-head tests for the new nms behavior.
  • Package version updated to 8.4.142.

🎯 Purpose & Impact

• ✅ Simpler API: One nms setting now controls head selection and export-time NMS across prediction, validation, tracking, benchmarking, and export.
• 🎯 Better default accuracy: YOLO26 now defaults to the one-to-many head with NMS, which generally provides higher accuracy than the NMS-free one-to-one head.
• ⚡ Optional lower-latency deployment: Use nms=False when you want NMS-free predictions and a simpler post-processing pipeline.
• 🔧 Easier migration: Existing end2end configurations continue to work through compatibility mapping, but new code should use nms.
• 📊 Fairer performance comparisons: Cross-format benchmark results are now more meaningful because all formats use matching square validation inputs.
• 🌍 Wider edge-device support: New and refined documentation makes deployment on Luxonis, Ambarella, DEEPX, Hailo, and other specialized platforms easier to follow.
• ⚠️ Export consideration: Some runtimes or quantized configurations cannot support NMS-free or embedded-NMS graphs and will fall back to native raw outputs. Check the target integration’s compatibility before deployment.

What's Changed

• Refine Ambarella integration documentation by @smohan-ambarella in #25465
• Add Luxonis OAK deployment documentation by @klemen1999 in #25452
• Document DEEPX SDK dependency installation by @Gornoka in #26067
• Use square validation inputs for cross-format benchmarks by @glenn-jocher in #26068
• Normalize augmentation overrides after restoring resume arguments by @glenn-jocher in #26069
• Unify end2end under nms=True|False|None by @glenn-jocher in #26066

New Contributors

• @smohan-ambarella made their first contribution in #25465
• @klemen1999 made their first contribution in #25452

Full Changelog: v8.4.141...v8.4.142

🌟 Summary

Ultralytics 8.4.141 updates Axelera support to Voyager SDK 1.8.0, adds batched inference, and improves model compatibility, dependency management, and CI reliability. 🚀

📊 Key Changes

• Axelera integration upgraded to Voyager SDK 1.8.0 🔧

  • Export and inference now use matching axelera-devkit and axelera-rt 1.8.0 packages.
  • Axelera export environments move to Python 3.13 and support a wider compatible PyTorch range.
  • Removed the previous NumPy version ceiling and improved SDK version checks before compiler loading.
  • Export now ensures the correct SDK environment tools are available on PATH and safely restores environment variables afterward.

• Batched Axelera inference is now supported 📦

  • Predictions with batch>1 are routed through the Voyager scheduler.
  • Outputs are returned in the original input order for both single- and multi-output models.
  • Models are still compiled for a single image, so batching improves workflow compatibility but does not necessarily increase end-to-end speed on one device.

• Axelera deployment requirements are documented more clearly 📚

  • Voyager SDK 1.8.0 requires the Metis kernel driver 1.6.2 or newer.
  • Models exported with an older SDK must be re-exported because SDK 1.8 uses a newer compiled model format.
  • YOLO26 semantic segmentation is documented as supported, while YOLO26 instance segmentation remains unavailable through the Ultralytics export command.

• Fixed legacy SPPF model reconstruction 🛠️

  • Pre-YOLO26 YAML configurations once risked rebuilding SPPF layers with the wrong activation behavior.
  • Legacy models now restore their original activation settings, preserving released model graphs and outputs.
  • Regression coverage now checks multiple YOLO configurations, including YOLO8, YOLO10, YOLO11, and YOLO26.

• OpenVINO dependencies are now opt-in 📦

  • OpenVINO and NNCF moved from export-base to a dedicated export-openvino extra.
  • The aggregate export extra still includes OpenVINO support.
  • This reduces unnecessary installations for users who only need the common export stack.

• Conda CI dependency resolution improved ⚡

  • CI now uses only conda-forge with strict channel priority and removes inherited default channels.
  • Installation is noninteractive and has a 20-minute timeout, preventing stalled dependency solves from consuming the full job.

🎯 Purpose & Impact

• Axelera users can use the latest SDK, run inference on image batches, and benefit from more reliable automatic dependency handling. Existing Axelera models may need to be re-exported after upgrading. ⚠️
• Deployment compatibility is improved, but users must keep the Metis driver aligned with Voyager SDK 1.8.0.
• Legacy Ultralytics models are safer to rebuild from YAML without silently changing SPPF behavior.
• General installations become lighter when OpenVINO is not required, while OpenVINO export remains available through its dedicated extra or the full export installation.
• Contributors and maintainers benefit from faster, more predictable Conda CI failures and clearer environment setup.
• The release increments the package version to 8.4.141, enabling downstream Ultralytics Platform updates to require the corrected Axelera exporter.

What's Changed

• Restore SPPF cv1 activation for pre-YOLO26 models by @Erickrk in #26040
• Fix Conda CI dependency resolution and install timeout by @glenn-jocher in #26064
• Move OpenVINO dependencies out of the base install extra by @Gornoka in #26038
• Update Axelera SDK to 1.8.0 and support batched inference by @glenn-jocher in #26065

New Contributors

• @Erickrk made their first contribution in #26040

Full Changelog: v8.4.140...v8.4.141

🌟 Summary

v8.4.140 improves training reliability and model-state preservation, led by a fix for grayscale TIFF datasets that could previously crash YOLO26 segmentation training.

📊 Key Changes

• 🖼️ Fixed grayscale TIFF channel handling (PR #26061, @glenn-jocher)
  Single-frame grayscale TIFF images now respect the requested color format. This prevents three-channel models from receiving one-channel batches and failing at the first convolution. Existing support for color TIFFs, multipage files, four-channel images, and multispectral stacking is preserved.

• 🧪 Expanded regression coverage for TIFF training
  The existing multichannel test now also trains, validates, predicts, and exports with grayscale TIFF data.

• 🧠 Preserved model state during inference and validation (PR #26062)
  Prediction and standalone validation now operate on independent model copies, preventing inference-time fusion, FP16 conversion, or end-to-end configuration changes from permanently modifying the caller’s model.

• 🚀 More reliable distributed training (PR #26050)
  DDP workers now receive the parent trainer’s prepared model, arguments, and callbacks. In-memory weight changes, custom class names, and application callbacks are therefore retained during multi-GPU training. The release adds cloudpickle to support this state transfer.

• 🎯 Improved weight loading and provenance (PR #26039)
  Loading weights from modules, checkpoint dictionaries, or files now preserves the correct training source and prefers EMA weights when available. Predictor caches are refreshed after model changes, reducing the risk of stale inference behavior.

• 🔧 Safer fusion detection and calibration (PR #26045)
  Fused pretrained weights now generate a warning when loaded into an unfused model. Fusion detection is more accurate for convolutional, reparameterized, and end-to-end YOLO26 components. Depth calibration also preserves a trainable, saveable model state.

• 📦 Safer exports (PR #26052)
  Exporting works on an isolated model copy before modifying names or head settings. This keeps the original model unchanged and avoids copying YOLOWorld’s large cached CLIP encoder unnecessarily.

• 🧬 Tuning now uses the caller’s loaded model (PR #26053)
  Model.tune() preserves weights that were loaded or modified in memory instead of rebuilding each tuning iteration from the original model path.

• 📈 Small dataset fractions no longer discard all data (PR #26049)
  Positive sampling fractions now retain at least one image, avoiding dataset-loading failures caused by rounding very small fractions down to zero. Explicit zero splits remain supported.

• ⚡ Faster single-image preprocessing (PR #25989)
  Single-frame NumPy inputs bypass an unnecessary stacking operation, improving preprocessing speed while retaining batched-input behavior.

• 🧹 Updated documentation and package version
  Inference and validation documentation now reflects the corrected model-state behavior, and the package version is bumped to 8.4.140.

🎯 Purpose & Impact

• ✅ Grayscale TIFF training jobs should run successfully instead of failing because of an image/model channel mismatch.
• 🛡️ Models are less likely to be unexpectedly altered by prediction, validation, calibration, export, or precision changes.
• 🏋️ Training and tuning become more faithful to the model the user actually loaded, including in-memory edits and custom weights.
• 🌐 Multi-GPU training is more consistent with single-GPU training, including support for custom callbacks and prepared model state.
• 📊 Small dataset experiments become more robust, especially when using fractional sampling.
• ⚡ Single-image inference receives a modest preprocessing speed improvement.
• 🔄 Users working with YOLO26, multispectral data, TIFF datasets, model export, or DDP training are likely to see the greatest benefits from this release.

What's Changed

• Keep the caller's depth model unfused across calibrate() and warn when pretrained weights are fused by @raimbekovm in #26045
• Keep a nonzero fraction from selecting zero images by @ahmet-f-gumustas in #26049
• Export a copy of the model before writing its head and names by @raimbekovm in #26052
• Seed tuning iterations from the model the caller loaded by @raimbekovm in #26053
• Hand DDP workers the module the parent trainer prepared by @raimbekovm in #26050
• Preserve loaded weights and checkpoint provenance for training by @raimbekovm in #26039
• Single frame special case by @jahsef in #25989
• Preserve inference model state and automatic DDP callbacks by @glenn-jocher in #26062
• Fix grayscale TIFF training channel mismatch by @glenn-jocher in #26061

Full Changelog: v8.4.139...v8.4.140

🌟 Summary

v8.4.139 improves training efficiency and reliability—especially by reducing validation memory usage—while refreshing model statistics, dataset handling, optimizer performance, and documentation. 🚀

📊 Key Changes

• Lower validation dataloader memory usage by @glenn-jocher:

  • Validation workers now prefetch 2 batches instead of 4.
  • Training loaders continue using a prefetch factor of 4.
  • This reduces queued validation data and shared-memory usage without slowing validation startup. 💾

• Faster Muon/MuSGD optimizer updates:

  • Updates are now grouped by column count and flattened according to their memory layout.
  • This avoids unnecessary tensor copies for channels-last convolution weights and preserves fused parameter updates.
  • The standalone Muon optimizer and related dead code were removed; MuSGD remains available. ⚡

• More reliable EMA checkpoints:

  • EMA attributes such as class names, class counts, strides, and class weights are now copied from the unwrapped model.
  • This fixes missing or stale metadata when using distributed training or torch.compile.
  • EMA is initialized after class weights are calculated, improving checkpoint consistency. ✅

• Correct semantic segmentation dataset detection:

  • A default masks/ directory is now recognized as PNG-mask semantic segmentation when no masks_dir entry is specified.
  • Prevents incorrect class counts and phantom background classes in affected datasets. 🎯

• Improved FP16 behavior documentation:

  • Documentation now clearly warns that quantize=16 can cast a retained PyTorch model in place.
  • A later FP32 call on the same model object may therefore use FP16-rounded weights.
  • This clarification applies to prediction and validation, including shared model objects. ⚠️

• Refreshed FLOPs and parameter reporting:

  • Published model statistics were synchronized with the current profiler.
  • Updated values cover YOLO26, YOLO11, YOLO12, YOLOv8, YOLOv5u, and other documented models.
  • Accuracy and speed benchmark values were not changed. 📊

• More robust and maintainable internals:

  • Batch-normalization fusion now uses PyTorch’s official fusion utilities.
  • SAM positional embeddings avoid nondeterministic cumulative-sum operations.
  • Temporary module aliases are now thread-safe and restore modified attributes correctly.
  • NumPy-to-tensor conversion can share CPU memory and perform a single device transfer.
  • ONNX export opset selection was simplified and made more compatible with current runtimes.

• Expanded task visibility and documentation:

  • Depth estimation is now listed consistently across the README, package description, docstrings, tutorials, and task detection messages.
  • Documentation validation now requires Zensical 0.0.58 or newer.
  • Several model benchmark tables and page frontmatter entries were corrected. 📚

🎯 Purpose & Impact

• Users training with limited system or shared memory should see fewer validation-memory spikes, particularly with many dataloader workers or large images. 🧠
• Training performance may improve for the default CUDA channels-last path because Muon updates avoid repeated copies and retain fused optimizer operations.
• Distributed and compiled training runs should produce more complete and dependable checkpoints through improved EMA metadata handling.
• Semantic segmentation workflows using a conventional root-level masks/ folder now behave as documented, avoiding incorrect output channels and class definitions.
• FP16 users should be aware that prediction or validation may modify a PyTorch model object in place; keeping separate model instances is recommended when switching between FP16 and FP32.
• Benchmark comparisons are more trustworthy because FLOPs and parameter counts now reflect current profiler behavior, although these corrected statistics may differ from earlier documentation.
• Depth estimation is easier to discover for new users, while existing model APIs and task support remain compatible. 🌊

What's Changed

• Document that a predict-time FP16 request casts the caller's PyTorch model in place by @raimbekovm in #26037
• Republish FLOPs measured with the current profiler by @raimbekovm in #26036
• Bucket Muon matrices by column count and flatten updates in memory order by @Y-T-G in #26035
• Fix semantic nc for the default masks/ folder by @laodouya in #26034
• Fix Zensical 0.0.58 docs validation by @glenn-jocher in #26041
• Replace hand-rolled torch helpers with their torch owners and drop dead code by @glenn-jocher in #26043
• Refresh EMA attributes from the unwrapped model by @glenn-jocher in #26047
• Add depth estimation to README intro paragraph by @raimbekovm in #26055
• Add depth estimation to task lists across README, docstrings and package description by @raimbekovm in #26056
• Reduce validation dataloader memory with prefetch 2 per worker by @glenn-jocher in #26057

Full Changelog: v8.4.138...v8.4.139

🌟 Summary

Ultralytics v8.4.138 is a stability-focused release that restores compatibility with older checkpoints, fixes YOLO-World loading, and improves training, inference, tracking, tuning, and documentation reliability. 🛠️

📊 Key Changes

• Legacy checkpoint loading fixed 🎯
  Restricted checkpoint loading now recognizes loss and assignment classes stored in checkpoints created before 8.4.95, including detection, classification, pose, segmentation, rotated-box, and keypoint-related components.

• YOLO-World loading fixed 🌍
  Corrects package loading issues affecting YOLO-World models and related checkpoints.

• MuSGD with channels_last no longer crashes ⚡
  Replaces an incompatible tensor flattening operation with a layout-safe one, allowing CUDA training with the MuSGD optimizer and automatic channels_last support to run correctly.

• SAM feature extraction optimized 🧠
  SAM, SAM2, and SAM3 image embeddings are computed in inference mode, preventing unnecessary autograd graphs from being retained and reducing memory overhead during repeated inference.

• Distributed training made more robust 🔧

  • Prevents DDP failures when a mini-batch contains no assigned targets.
  • Ensures detection, pose, OBB, and related model branches remain connected to the training graph even when there are no positive samples.

• Tuning results corrected 📈
  Multi-dataset tuning now preserves dataset names and iteration order across distributed workers. Failed datasets are recorded with zero metrics instead of incomplete results, keeping tuning histories and fitness plots consistent.

• Classification inference preprocessing improved 🚀
  More preprocessing work is moved to the inference device and performed in batches, which can reduce CPU overhead and improve classification throughput.

• BoT-SORT tracking made faster 🏃
  Global motion compensation now caps corner detection at 400 points instead of 1,000, reducing optical-flow computation while retaining sufficient information for motion estimation.

• Precision and quantization documentation clarified 📚
  Documentation now explains that quantize may select or request different runtime precisions depending on the export format. This avoids implying that every backend supports the same FP16, FP32, or INT8 behavior.

• Documentation quality updates ✨
  Markdown tables were consistently formatted, the OBB navigation label was cleaned up, the TrackZone video was updated, and the YOLO26 CPU speed comparison now clearly identifies its YOLO26n-versus-YOLO11n ONNX baseline and hardware.

🎯 Purpose & Impact

• More users can load existing models without retraining, especially those using checkpoints created with older Ultralytics versions. ✅
• YOLO-World workflows become more dependable for users loading supported models and checkpoints.
• Platform and cloud GPU training jobs are less likely to fail, particularly MuSGD runs using CUDA and channels_last.
• Distributed training and hyperparameter tuning produce more trustworthy results, even with empty-target batches or failed datasets.
• SAM inference can use less memory, which is helpful for long-running applications and repeated image embedding extraction.
• Tracking pipelines may run with lower latency, especially in applications where BoT-SORT global motion compensation is a significant cost.
• Users should review the updated quantization guidance before assuming that quantize=16 or quantize=32 produces the same computation precision across PyTorch, ONNX, OpenVINO, NCNN, MNN, Triton, and other backends. 🔍

What's Changed

• Optimize SAM image feature extraction by @glenn-jocher in #26011
• Remove inconsistent OBB Dataset in the docs navigation bar by @Laughing-q in #26019
• Align Markdown tables inside MkDocs tabs and admonitions by @Laughing-q in #26020
• Fix distributed tuning fitness dataset labels by @glenn-jocher in #26024
• Add https://youtu.be/Zikt3a7PngE to docs by @RizwanMunawar in #26005
• Name the YOLO26n vs YOLO11n baseline for the 43% CPU figure by @raimbekovm in #26025
• Halve BoT-SORT global motion compensation cost by capping corner detection at 400 by @JESUSROYETH in #26008
• Fix MuSGD crash with channels_last: use reshape for non-contiguous grads by @synml in #26013
• Fix DDP unused-gradient crash when a mini-batch has no assigned targets by @JESUSROYETH in #26012
• Document what precision quantize selects for each format by @raimbekovm in #26022
• Move classification predictor preprocessing to the inference device by @JESUSROYETH in #26023
• Fix tuner metrics for failed datasets by @glenn-jocher in #26028
• Remove redundant regression scaffolding by @glenn-jocher in #26029
• Fix legacy checkpoint and YOLO-World loading by @glenn-jocher in #26032

Full Changelog: v8.4.137...v8.4.138

🌟 Summary

v8.4.137 automatically enables the faster channels-last memory layout for CUDA training on PyTorch 1.11+, improving GPU training performance while preserving clear opt-out and compatibility options. 🚀

📊 Key Changes

• Automatic channels-last training: The existing channels_last=None setting now automatically uses the NHWC memory format for CUDA training with PyTorch 1.11 and newer.
• Explicit control remains available:
  • channels_last=None: Automatically selects channels-last when supported.
  • channels_last=False: Explicitly keeps the traditional NCHW format.
  • channels_last=True: Explicitly requests channels-last, preserving previous behavior.
• Safe compatibility behavior: PyTorch 1.10 and older, CPU, and MPS training continue using NCHW by default.
• Improved resume handling: The channels_last setting is now included among the training options that can be updated when resuming a run.
• Documentation updates: Training guides and argument references now describe the automatic CUDA behavior.
• No model architecture changes: This release focuses on training performance and memory layout rather than changing model structure or outputs.

🎯 Purpose & Impact

• ⚡ Potentially faster CUDA training: Channels-last can improve convolution performance on compatible GPUs, particularly modern Tensor Core hardware.
• 🧠 Better YOLO26 training defaults: Users no longer need to manually enable the optimization when using a supported PyTorch and CUDA environment.
• 🛡️ Reduced compatibility risk: Automatic activation begins at PyTorch 1.11, the first validated version that avoids known channels-last failures in YOLO26 training.
• 🔧 Full user control: Workloads that require the traditional layout can disable the optimization with channels_last=False.
• 🌍 Broad validation: The change was tested across CUDA 11.1–13.2, PyTorch 1.8–2.12, Python 3.8–3.13, and a wide range of modern NVIDIA GPUs.
• 📦 Minimal implementation impact: The behavior is selected during trainer setup without adding new arguments, persistent state, helper utilities, or GPU-specific allowlists.

What's Changed

• Auto-enable channels-last CUDA training by @glenn-jocher in #26007

Full Changelog: v8.4.136...v8.4.137

🌟 Summary

Version 8.4.136 improves hyperparameter tuning, inference performance, backend compatibility, and data handling—making YOLO workflows more reliable and efficient. 🚀

📊 Key Changes

• 🎯 Smarter hyperparameter tuning — Current PR #25996 by @glenn-jocher

  • Keeps the existing Gaussian search behavior unchanged during the first 30 completed trials.
  • Learns relationships between promising hyperparameters by analyzing the best-performing results.
  • Uses confidence-weighted, correlated mutations only when enough elite trial data is available.
  • Reflects correlated proposals at search boundaries to avoid repeatedly clipping values.
  • Benchmarking on basketball-hoop detection reported a new best fitness of 0.60576, outperforming the tested Ray Tune and previous custom Tuner configurations.
  • Updated the hyperparameter tuning documentation.

• 🧠 More robust channels-last inference

  • AutoBackend now owns memory-layout selection during backend construction, avoiding duplicated or unsafe conversions.
  • Automatic channels-last selection is available for supported Linux and Windows x86 CPU environments using PyTorch 1.13 or newer with oneDNN.
  • CUDA support remains available, while ARM64, MPS, older PyTorch versions, and exported backends retain their existing behavior.
  • Fixed compatibility issues affecting PyTorch 1.9 and JetPack 6 systems.

• ⚡ Faster image preprocessing

  • PIL and NumPy image inputs now use more efficient OpenCV color conversions.
  • Avoids unnecessary image copies while preserving correct channel order and contiguous memory layout.

• 🔍 More reliable prediction filtering

  • Fixed the CLI classes filter for YOLOE and World models when class IDs are supplied numerically.
  • Text-based class prompts continue to work as before.

• 📷 Improved image and visualization handling

  • TIFF loading now respects uppercase extensions and grayscale flags, preserving multispectral image channels correctly.
  • Pose visualization now scales keypoint coordinates without incorrectly scaling confidence values.
  • Matplotlib backend restoration is safer when the originally configured backend is unavailable.

• 🏃 Tracking and distributed tuning improvements

  • BoT-SORT sparse optical-flow tracking avoids an unnecessary per-pixel grid allocation, reducing overhead on large frames.
  • MongoDB-based distributed tuning now assigns iteration IDs atomically, preventing duplicate trial numbers from concurrent workers.

• 🧪 Better validation and project maintenance

  • Dataset YAML files now receive early type validation with clearer error messages.
  • Added CI and PyPI publishing status for the Ultralytics SDK repository.
  • Updated the package version to 8.4.136.

🎯 Purpose & Impact

• Better tuning results: The Tuner can discover useful relationships between hyperparameters instead of treating every parameter independently, potentially improving final model quality with fewer wasted trials. 📈
• Safer inference across platforms: Backend construction now handles memory formats and retained tensors more consistently, reducing regressions on older PyTorch versions, ARM64 devices, and JetPack environments.
• Faster predictions: Common PIL and NumPy input paths require fewer copies and more efficient conversions, which can improve throughput in image-heavy applications.
• More predictable CLI behavior: Class filtering now works consistently across standard, YOLOE, and World models.
• Improved dataset reliability: Invalid YAML field types are reported earlier, making dataset configuration errors easier to diagnose.
• No architecture changes: This release does not introduce a new model architecture; its primary benefits are improved tuning, compatibility, performance, and correctness.

What's Changed

• Add Platform SDK CI status by @glenn-jocher in #25997
• Fix legacy torch predict regressions from default CPU channels-last by @onuralpszr in #25999
• Own channels-last layout during backend construction by @glenn-jocher in #26001
• Restore automatic CPU channels-last selection by @glenn-jocher in #26002
• Speed up PIL and ndarray prediction sources with cv2 color conversion by @JESUSROYETH in #25988
• Fix plot_images scaling pose keypoint confidence with coordinates by @aleehydar in #24432
• Fix imread TIFF branch ignoring flags and extension case by @lszSamLin in #25986
• Fix CLI classes filter being dropped for world and yoloe models by @chinazhangchao in #25882
• Skip OpenCV's per-pixel grid in BoT-SORT motion compensation by @JESUSROYETH in #25995
• Add type checking for dataset YAML by @Y-T-G in #21905
• Fix Matplotlib error when restoring original backend by @Y-T-G in #21863
• Fix concurrent MongoDB tuner iteration IDs by @glenn-jocher in #26000
• Fix PyTorch 1.9 backend initialization by @glenn-jocher in #26003
• Improve Tuner search with confidence-weighted covariance by @glenn-jocher in #25996

New Contributors

• @chinazhangchao made their first contribution in #25882
• @aleehydar made their first contribution in #24432
• @lszSamLin made their first contribution in #25986

Full Changelog: v8.4.135...v8.4.136

🌟 Summary

v8.4.135 improves detection reliability by adapting max_det to dataset object counts and standardizing dataset fraction handling.

📊 Key Changes

• 🚀 Smarter max_det selection for detection, segmentation, pose, and OBB tasks

  • Training and validation now inspect the largest number of labeled objects found in a single image.
  • If the default max_det is too low, it is automatically increased to match the observed dataset maximum.
  • User-specified max_det values are preserved, but a warning is shown when they may limit validation recall.
  • The resolved value is propagated to native end-to-end model heads before validation, improving consistency for NMS-free models.

• ⚠️ Clearer warnings for object-count mismatches

  • Users are notified when images contain more objects than max_det allows.
  • Warnings explain that a low limit can cap recall and produce misleading validation metrics.
  • Increasing max_det may increase validation cost, and cannot exceed the model or export format’s own capacity.

• 📏 Consistent fraction boundary behavior

  • fraction=1 and fraction=1.0 now both mean “use the full dataset.”
  • Integers greater than 1 continue to represent an image count.
  • 0 and 0.0 remain available for skipping an optional test split.
  • Training and validation splits must still contain at least one image.
  • Boolean values such as fraction=True are now rejected instead of being interpreted ambiguously.

• 📚 Documentation and validation updates

  • Training, export, and cloud-training documentation now describe the normalized fraction semantics.
  • Additional tests cover configuration validation, dataset conversion, concatenated datasets, training pipelines, and end-to-end detection behavior.

🎯 Purpose & Impact

• ✅ More trustworthy validation: Large-object-count images are less likely to be truncated by an unnoticed default limit.
• 📈 Better recall measurement: Automatically matching max_det to observed data helps prevent artificially low validation recall.
• 🧩 More predictable configuration: Dataset behavior no longer depends on whether a serializer writes 1 as an integer or 1.0 as a float.
• 🛠️ Safer user overrides: Custom max_det settings continue to work, with warnings when they may restrict results.
• ⚡ Potential performance trade-off: A higher detection limit can increase validation and inference post-processing cost, while model or deployment-format limits may still cap the maximum number of predictions.

What's Changed

• Normalize fraction boundary values by @glenn-jocher in #25994
• Respect dataset object counts when selecting max_det by @glenn-jocher in #25993

Full Changelog: v8.4.134...v8.4.135

🌟 Summary

v8.4.134 makes crowded-object training more resilient to GPU memory limits and increases the default hyperparameter search budget from 10 to 300 trials. 🚀

📊 Key Changes

• Faster TaskAlignedAssigner OOM recovery 🧠

  • When GPU memory runs out during target assignment, Ultralytics now retries the work one image at a time on the GPU instead of moving the entire operation to the CPU.
  • The model’s forward-pass batch size remains unchanged.
  • A warning is shown only once per training run and reports the assignment dimensions.

• Lower memory usage during target assignment 💾

  • Removes unused trailing ground-truth padding before retrying.
  • Preallocates retry outputs and reuses dense metric buffers.
  • Keeps candidate masks in a compact integer format.
  • Computes box metrics only for valid anchor/ground-truth pairs.
  • Uses more memory-efficient in-place operations where possible.

• Improved geometric and assignment processing ⚙️

  • Streamlines point-in-box checks and overlapping-ground-truth resolution.
  • Preserves compatible output behavior while reducing temporary tensor allocations.

• Hyperparameter tuning now defaults to 300 trials 🔍

  • The built-in tuner, direct Tuner usage, and Ray Tune all now use 300 trials by default, up from 10.
  • Documentation has been updated to reflect the shared default.

• Version update 📦

  • Package version updated to 8.4.134.

🎯 Purpose & Impact

• Much better recovery from GPU out-of-memory errors ✅
  Large or highly crowded batches can continue training without falling back to a very slow full-CPU assignment. In the reported xView benchmark, optimized single-image GPU assignment used about 1.56 GB of peak memory and completed in 0.376 seconds, compared with 194.86 seconds and 85.62 GB for the previous GPU-to-CPU fallback.

• More practical training on dense datasets 🏙️
  The changes are particularly valuable for aerial imagery, crowd analysis, and other datasets containing many objects per image.

• No need to reduce the model’s forward batch size after recovery 📈
  The fallback is isolated to target assignment, helping retain the intended training configuration while handling temporary memory pressure.

• More effective automatic hyperparameter tuning 🎯
  A 300-trial default gives the tuner substantially more opportunities to explore configurations and find stronger settings, especially for the broad YOLO26 search space.

• Higher tuning cost unless overridden ⏱️
  Users who rely on defaults should expect tuning jobs to run considerably longer and consume more compute. Smaller runs can still be requested explicitly when time or budget is limited.

What's Changed

• Default hyperparameter tuning to 300 trials by @glenn-jocher in #25987
• Reduce TaskAlignedAssigner OOM recovery cost by @glenn-jocher in #25990

Full Changelog: v8.4.133...v8.4.134

🌟 Summary

Ultralytics 8.4.133 improves hyperparameter tuning convergence, speeds up inference preprocessing, expands detection metrics, and simplifies edge-device setup. 🚀

📊 Key Changes

• Smarter hyperparameter tuning — PR #25984 by @glenn-jocher

  • Replaces coordinate-by-coordinate crossover with fitness-weighted selection of complete, high-performing configurations.
  • Preserves useful relationships between hyperparameters instead of mixing them independently.
  • Mutates approximately half of the parameters in normalized search-space coordinates, allowing parameters that start at zero—such as degrees or shear—to evolve more effectively.
  • Gradually reduces mutation size when tuning stops finding better results, encouraging refinement after broad exploration.
  • Prevents duplicate candidates after clipping, rounding, or integer conversion, including small and discrete search spaces.
  • Ray Tune now defaults to Optuna multivariate TPE, with parallel-aware suggestions rather than independent random search.

• Faster predictor preprocessing — PR #25982 by @jahsef ⚡

  • Moves image channel reordering and tensor-contiguity operations from CPU-side NumPy processing to the inference device.
  • Preserves output values while reducing unnecessary CPU copies.
  • Reported benchmarks show approximately 2.2–3.1× faster preprocessing on an RTX 5080, with additional gains on CPU.

• Automatic channels-last CPU inference — PR #25983 by @JESUSROYETH

  • Enables channels-last memory layout automatically for native PyTorch inference and standalone validation on supported x86 Linux and Windows CPUs with oneDNN.
  • Keeps training defaults and unsupported platforms unchanged.
  • Explicit channels_last=True remains available for supported CPU and CUDA paths.
  • Saved models are converted back to a safe contiguous format and stale EMA data is cleared to improve compatibility.

• More accurate INT8 calibration subsets — PR #25978 by @JESUSROYETH

  • Fixes fraction handling during classification and detection INT8 export calibration.
  • Scalar fractions now apply directly to the selected calibration split, while list-based fractions retain train/validation/test behavior.
  • Prevents exports from unintentionally calibrating on an entire dataset when only a subset was requested.

• Size-specific mAP for custom detection datasets — PR #25981 by @fcakyon 📈

  • Custom detection datasets can now report small-, medium-, and large-object mAP when using save_json=True.
  • Builds temporary COCO-format annotations internally while preserving existing native metrics and prediction files.
  • Applies consistently during training validation, final-model validation, and standalone validation.

• Simpler edge-device installation

  • Raspberry Pi, Jetson, DGX Spark, DeepStream, and related guides now install the base ultralytics package instead of the larger [export] extra.
  • Export dependencies are installed automatically when an export is requested, reducing installation size and dependency conflicts.

• Improved Weights & Biases artifact control — PR #25985 by @glenn-jocher

  • W&B model artifact uploads now follow the existing training save argument.
  • save=False skips uploading the best checkpoint while retaining metrics and plots.
  • Default behavior remains unchanged with save=True.

• Package update

  • Version bumped to 8.4.133.

🎯 Purpose & Impact

• Better tuning results: Hyperparameter searches are more likely to preserve successful configurations, explore meaningful alternatives, and avoid wasting trials on duplicates. 🎯
• Faster inference: Device-side preprocessing can reduce latency, particularly for batched inference and CPU-bound pipelines.
• Broader performance optimization: Supported x86 CPU users may benefit from channels-last inference without changing their existing commands.
• More reliable model export: INT8 calibration now honors requested dataset fractions, improving calibration speed and reducing unexpected resource usage.
• Richer evaluation: Custom detection datasets can now receive object-size performance breakdowns similar to COCO evaluations.
• Easier edge deployment: Base installations are smaller and less prone to dependency conflicts, while export workflows remain available when needed.
• More control over experiment storage: W&B users can keep experiment tracking lightweight by disabling checkpoint saving with the standard save setting.

What's Changed

• Allow skipping W&B model artifacts by @fcakyon in #25979
• Remove [export] from edge-device install guides by @Y-T-G in #25977
• Accelerate predictor preprocessing on inference devices by @jahsef in #25982
• Fix INT8 export calibration fractions by @JESUSROYETH in #25978
• Use save argument for W&B model artifacts by @glenn-jocher in #25985
• Compute size-specific mAP for custom detection datasets by @fcakyon in #25981
• Enable channels-last by default for x86 CPU inference by @JESUSROYETH in #25983
• Improve hyperparameter Tuner mutation convergence by @glenn-jocher in #25984

Full Changelog: v8.4.132...v8.4.133

🌟 Summary

🚀 Ultralytics v8.4.132 improves dataset efficiency, hardware compatibility, export workflows, and experiment-tracking documentation, with the headline feature being finer control over test-split downloads.

📊 Key Changes

• 🎯 Test-split control with fraction (PR #25966 — @fcakyon)

  • List-based fraction now supports train, validation, and test splits.
  • Use a third value to limit test data or set it to 0 to skip test-image downloads entirely.
  • Existing two-value lists remain compatible and continue using the full test split.
  • The same split-selection behavior is shared across training, validation, dataset conversion, and export calibration.

• ⚡ More efficient NDJSON workflows

  • Platform NDJSON runs no longer need to download unused test images, which can significantly reduce transfer time, storage use, and bandwidth during multi-dataset sweeps.

• 🩹 Corrected NMS and validation on Ascend NPU

  • NMS now uses the optimized torchvision path on supported Ascend NPU setups while continuing to avoid unsupported XPU behavior.
  • Fixes repeated NMS time-limit warnings and prevents incomplete batches from producing incorrect mAP results.

• 📚 Improved YOLO26 end-to-end and export guidance

  • Documentation now more accurately describes NMS-free detection, segmentation, pose, and OBB models.
  • Clarifies export-format fallbacks, quantization limitations, Hailo behavior, detection limits, and expected accuracy trade-offs.
  • Updates performance claims and explains when custom post-processing or NMS is still required.

• 🔄 Broader support for exported non-YOLO models

  • Documents loading generic exported models through YOLO() when task and imgsz are supplied explicitly.
  • Adds guidance for using format-specific backend classes and explains input-layout, autograd, and post-processing differences.
  • Corrects TensorFlow SavedModel, LiteRT, and numerical-parity documentation.

• 🧪 Better classification subset sampling

  • Fraction-based sampling for local classification datasets now selects images across classes instead of taking a class-ordered prefix.
  • Produces more representative subsets for quick experiments and validation.

• 📈 Updated experiment-tracking integrations

  • Ray Tune now uses its public context and reporting APIs.
  • ClearML, Comet, TensorBoard, MLflow, Ray Tune, and W&B instructions now better match current callback behavior.
  • Neptune integration and settings were retired following the service shutdown.
  • MLflow environment-variable handling is more flexible, and TensorBoard setup is clearly documented as optional.

• 🧩 Additional training and export corrections

  • RT-DETR documentation recommends disabling deterministic mode on CUDA, since its attention operation cannot provide fully deterministic backward training.
  • Copy-Paste augmentation is correctly documented for both segment and OBB tasks.
  • Classification color-jitter settings are clarified when automatic augmentation is disabled.
  • IMX exports no longer reject end-to-end models at the exporter entry point.
  • CLA workflow matching now accepts signature comments with surrounding whitespace or text.

🎯 Purpose & Impact

• 💾 Lower data-transfer costs: Users running Platform or NDJSON workflows can avoid downloading test data they do not need.
• ⏱️ Faster experimentation: Smaller, representative dataset subsets make tuning and iteration quicker without changing existing two-item fraction usage.
• 📊 More trustworthy metrics: Ascend NPU validation now processes complete batches, improving the reliability of reported mAP.
• 🚀 Easier deployment: Clearer YOLO26 export guidance helps users understand when NMS-free inference works and when a traditional NMS pipeline is necessary.
• 🌍 Better hardware coverage: NPU and export-path fixes improve compatibility across specialized accelerators and deployment formats.
• 🧭 Smoother integrations: Updated tracking documentation reduces setup confusion, while users migrating from Neptune should select an active integration such as ClearML, Comet, MLflow, TensorBoard, W&B, or the Ultralytics Platform.

What's Changed

• Match CLA signature comments by substring in the workflow gate by @glenn-jocher in #25970
• Fix NMS time-limit truncation and wrong mAP on Ascend NPU by @idwenhui in #25963
• Update experiment tracking callbacks and docs by @glenn-jocher in #25972
• Add https://youtu.be/rMllxg8ZLs8 to docs by @RizwanMunawar in #25975
• Document running non-YOLO exports through Ultralytics and correct parity claims by @raimbekovm in #25974
• Correct end-to-end detection guide claims and export fallback coverage by @raimbekovm in #25971
• Correct copy-paste task attribution and the export conf default claim by @raimbekovm in #25969
• Fix classification fraction sampling by @JESUSROYETH in #25968
• Recommend deterministic=False for RT-DETR training by @artest08 in #25965
• Extend fraction to the test split by @fcakyon in #25966

New Contributors

• @idwenhui made their first contribution in #25963

Full Changelog: v8.4.131...v8.4.132

🌟 Summary

Ultralytics v8.4.131 adds Apple Core AI export and inference support for YOLO26, alongside important validation, training, model-configuration, and documentation improvements. 🚀

📊 Key Changes

• 🍎 Apple Core AI export and inference

  • Export models with model.export(format="coreai") or the equivalent CLI command.
  • Creates Apple’s .aimodel asset format, which can be loaded again with YOLO("yolo26n.aimodel").
  • Supports FP32 and optional FP16 export through the new export-coreai dependency group.
  • Adds a dedicated Core AI backend, metadata handling, API references, export-table support, and continuous macOS CI coverage.
  • Supports YOLO26 models on Apple silicon with macOS 26 or later; exported assets target iOS 27 and macOS 27.
  • Core AI export currently has important limitations: fixed input size, no dynamic shapes or NMS export, and no support in the Ultralytics iOS or Flutter SDKs yet.

• ⚡ Core AI deployment options

  • YOLO26’s end-to-end head is exported by default, returning finished detections directly.
  • Exporting with end2end=False produces raw predictions and can significantly reduce inference latency when post-processing is handled on the host.
  • Core AI export includes model metadata such as class names, stride, and task information inside the .aimodel asset.

• ✅ More reliable validation with split=train

  • Validation now consistently uses the unaugmented validation pipeline instead of accidentally applying training augmentations such as Mosaic, MixUp, and Random Perspective.
  • Dataset fractions are now selected according to the requested split.
  • This fixes crashes and unreliable metrics for detection, segmentation, OBB, RT-DETR, and YOLOE validation workflows.

• 🧮 Correct YOLO26 loss terminology

  • Documentation and logging now distinguish YOLO26’s l1_loss from dfl_loss used by models with distribution-based box regression.
  • Training guides, default configuration comments, tuning tables, experiment trackers, and tutorial output have been updated accordingly.

• 🎯 Improved model configuration handling

  • Model-scale overrides in parse_model now match exact scale letters, preventing unscaled or dictionary-based configurations from taking the wrong architecture branch.
  • C3k2 configurations without an explicitly provided optional argument no longer fail for medium, large, or extra-large variants.
  • The architecture guide now explains these scale-dependent behaviors more accurately.

• 🔤 YOLOE class reordering fixes

  • YOLOE.set_classes() now recognizes class-order changes and regenerates prompt embeddings when necessary.
  • Reordering classes therefore updates class IDs and names correctly instead of being treated as a no-op.

• ⚖️ Training robustness improvements

  • Class weights are now preserved on the underlying model during DDP training and continue to target the student model correctly during knowledge distillation.
  • Fine-tuning guidance now recommends non-zero warmup while clarifying that the full three-epoch default is not always necessary.
  • Documentation now accurately describes automatic optimizer selection and module-name-based layer freezing.

• 📟 Better progress bars in notebooks and narrow terminals

  • TQDM output no longer disappears in zero-width pseudo-terminals such as those used by Colab.
  • Notebook output is allowed to scroll naturally, while truncated terminal lines now show an ellipsis instead of being silently cut off.

• 📚 Documentation and presentation updates

  • Corrects documented YOLOE and YOLOv5 run paths to match actual increment_path behavior.
  • Adds the missing OBB task header image.
  • Expands and updates Apple Core AI integration guidance, including deployment limitations and Core ML recommendations.

🎯 Purpose & Impact

• Apple developers gain a new native deployment path for YOLO26 models on the latest Apple silicon platforms, potentially improving on-device inference integration and performance. 🍏
• Core ML remains the better choice for broader compatibility, including older Apple operating systems and the Ultralytics iOS or Flutter SDKs. Core AI is currently best suited to experimental or platform-specific deployments.
• Validation results become more trustworthy, especially when evaluating training data with split=train, because training-time augmentation is no longer applied accidentally.
• YOLO26 users receive clearer training feedback, avoiding confusion between L1 box-distance loss and DFL metrics.
• Custom and scaled model definitions are more dependable, reducing silent architecture mismatches and configuration-related errors.
• YOLOE class prompts are safer to update, particularly in applications that dynamically change class ordering.
• Distributed training and notebook workflows become more reliable, with class weights preserved correctly and progress output rendered consistently.

What's Changed

• Match documented run paths to each product's own increment_path by @raimbekovm in #25942
• Match model scale letters exactly in parse_model and correct the architecture guide by @raimbekovm in #25940
• Add the OBB task header image by @raimbekovm in #25947
• Correct the dfl loss description for DFL-free YOLO26 by @raimbekovm in #25941
• Correct training-argument documentation that disagrees with the trainer by @raimbekovm in #25957
• Preserve class_weights during DDP training by @Laughing-q in #25959
• Apply YOLOE class updates when only the order changes by @aswanth-07 in #25956
• Fit the TQDM bar on zero-width and notebook terminals by @Y-T-G in #25958
• Keep validation unaugmented on split=train and pick the fraction by split by @onuralpszr in #25962
• Add Apple Core AI export by @john-rocky in #25926

Full Changelog: v8.4.130...v8.4.131

🌟 Summary

Version v8.4.130 makes dataset subset selection far more flexible and efficient, while improving tuning, tracking guidance, and dataset metadata. 🚀

📊 Key Changes

• Count-based dataset limits 🎯

  • fraction now accepts a positive image count, such as fraction=1000, to train on exactly 1,000 images.
  • Use fraction=[1000, 100] to limit the training and validation splits independently.
  • Existing decimal ratio behavior remains unchanged, so fraction=0.1 still uses 10% of the dataset.
  • Integer 1 means one image, while float 1.0 means the complete split.
  • Supports YOLO, RTDETR, classification, validation, and INT8 calibration workflows.

• More efficient NDJSON and Platform dataset downloads ⚡

  • Count-based subsets are selected before images are downloaded.
  • NDJSON records are selected deterministically, helping repeated runs use the same images.
  • This avoids downloading an entire dataset when only a fixed-size subset is needed.

• Improved hyperparameter tuning 🧠

  • Model.tune() now defaults to AdamW unless another optimizer is explicitly selected.
  • This ensures tuning parameters such as learning rate and momentum actually affect training instead of being ignored by automatic optimizer selection.
  • MongoDB-based tuning now uses safer atomic coordination, preventing multiple workers from incorrectly claiming the default configuration.

• Clearer tuning fitness plots 📈

  • tune_fitness.png now shows overall fitness progression, the best result achieved so far, and initial-versus-best fitness for each dataset.
  • The new layout is easier to interpret, especially for multi-dataset tuning runs.

• Expanded and clarified tracking documentation 🎥

  • Documentation now lists six built-in trackers: TrackTrack, BoT-SORT, ByteTrack, OC-SORT, Deep OC-SORT, and FastTracker.
  • TrackTrack is documented as the default tracker, with optional ReID and camera-motion compensation.
  • Tracking guidance now more clearly explains confidence thresholds, low-confidence recovery, custom ReID models, and task-specific behavior for segmentation, pose, and OBB models.
  • Tracker-specific training is clarified: users train a detection, segmentation, pose, or OBB model, then apply tracking during inference.

• More complete dataset license metadata 📚

  • Added or corrected license information for MNIST, Global Wheat2020, PASCAL VOC, KITTI, and official depth datasets.
  • Depth8 and SUN RGB-D are now explicitly marked as having no specified source license where applicable.
  • Export documentation across ONNX, TensorRT, OpenVINO, LiteRT, Hailo, QNN, Rockchip, and other formats now reflects the expanded fraction behavior.

🎯 Purpose & Impact

• Faster experimentation: Quickly train or calibrate on a known number of images without creating duplicate dataset copies.
• Lower storage and bandwidth usage: Platform NDJSON datasets no longer need to download every image before applying a count-based limit.
• More reliable tuning: AdamW makes the default tuning search spaces effective, while MongoDB coordination avoids duplicate baseline trials in concurrent runs.
• Better reproducibility: Deterministic NDJSON subset selection makes repeated experiments more consistent.
• Improved deployment workflows: Fixed-size calibration subsets are now easier to use across supported export formats, helping reduce INT8 calibration time.
• Clearer tracking decisions: Users can more easily choose a tracker and understand the trade-offs between speed, ReID, camera-motion compensation, and occlusion handling.
• No major model architecture changes: This release primarily improves data handling, tuning reliability, tracking usability, and documentation rather than introducing a new model family.

What's Changed

• Fix concurrent MongoDB tuner default claims by @glenn-jocher in #25939
• Fix Tune optimizer default by @glenn-jocher in #25945
• Improve tuning fitness plot by @glenn-jocher in #25946
• Add depth dataset license metadata by @glenn-jocher in #25948
• Fix Depth8 licensing and tracker docs by @glenn-jocher in #25950
• Enable fraction to limit dataset by image counts by @fcakyon in #25951

Full Changelog: v8.4.129...v8.4.130

🌟 Summary

v8.4.129 improves multi-dataset hyperparameter tuning, training precision, model export acceleration, data loading, and platform reliability—without introducing a new model architecture. 🚀

📊 Key Changes

• Multi-dataset tuning is now managed by MultiTrainer (PR #25937, @glenn-jocher):

  • Removes duplicated dataset orchestration from the tuner.
  • Runs each dataset training job in an isolated YOLO CLI subprocess.
  • Preserves YOLOWorld and YOLOE checkpoint filename handling.
  • Records individual dataset metrics, macro-mean metrics, and cleanup paths.
  • Improves distributed MongoDB tuning by safely assigning defaults and identifying winning runs by their result paths instead of worker-local indexes.
  • Validated with single-dataset and five-dataset MongoDB tuning runs. ✅

• Added BF16 mixed-precision training (PR #25931, @artest08):

  • amp now accepts True, False, "fp16", "bf16", and "fp32".
  • BF16 uses less memory than FP32 while offering greater numerical range than FP16.
  • Gradient scaling is correctly disabled for BF16.
  • Supported CUDA hardware is required for native BF16 training.

• Improved YOLO26 LiteRT exports for GPU delegates (PR #25914, @Y-T-G):

  • Reworked detection-head indexing and gathering to use operations better supported by GPU accelerators.
  • Helps keep more of the end-to-end, NMS-free detection head on the GPU instead of falling back to the CPU.
  • Also centralizes export-specific behavior outside the main detection head implementation.

• Faster image and FastSAM preprocessing (PRs #25935 and #25938, @JESUSROYETH):

  • Large batches of regular images can now be decoded in parallel while preserving input order.
  • FastSAM CLIP crop preprocessing is parallelized for sufficiently large CUDA workloads.
  • Small batches, unsupported image formats, CPU, and MPS paths retain the safer serial behavior.

• More consistent ONNX CPU benchmarking (PR #23924, @Laughing-q):

  • ONNX profiling now uses the shared ONNXBackend.
  • Adds configurable ONNX Runtime session options and multi-input model support.
  • Benchmarks now follow the same backend execution path used during inference.

• Stronger export and validation coverage:

  • TensorRT tests now validate exported engines on task-specific datasets, not only through inference.
  • CoreML and LiteRT export helpers are documented in the API reference.
  • Detection postprocessing is shared across Detect, Segment, Pose, OBB, and related heads.

• Improved Windows image compatibility (PR #21070, @Laughing-q):

  • Added imread_unicode for image paths containing non-ASCII characters.
  • Preserves native OpenCV grayscale behavior and simplifies semantic-mask handling.

• More reliable progress reporting for Ultralytics Platform (PR #25905, @Y-T-G):

  • Progress-bar redraws are now transmitted as live state rather than ordinary log lines.
  • Reduces duplicated or cluttered logs and gives consumers a cleaner progress contract.

• Documentation and training guidance corrections:

  • Clarifies YOLO26’s DFL-free l1_loss, pretrained-weight behavior, AutoBatch rules, AMP behavior, freezing, fine-tuning, K-Fold workflows, model YAML construction, and tuning output paths.
  • Adds guidance for class-name-based head weight remapping during fine-tuning.
  • Documents BF16 settings and supported distillation task limitations.

• Build reliability improvements for Jetson (PR #25930, @glenn-jocher):

  • Disk cleanup is skipped for JetPack Docker builds so native Jetson builds retain sufficient runner swap space.

🎯 Purpose & Impact

• More dependable distributed tuning: Multi-dataset experiments are now simpler to maintain, better isolated, and easier to analyze because results and output paths are tracked per dataset. MongoDB workers are also less likely to race when initializing defaults. 📈

• More training options: Users with compatible CUDA hardware can choose BF16 for a practical balance of speed, memory use, and numerical stability. Existing FP16 and FP32 behavior remains available.

• Faster edge deployment: LiteRT exports of YOLO26 end-to-end models should make better use of WebGPU and other GPU delegates, reducing unnecessary CPU fallback and potentially improving inference latency. ⚡

• Better throughput for large workloads: Parallel image decoding and FastSAM preprocessing can reduce time spent waiting for CPU preprocessing, especially with large batches or many candidate crops.

• Improved compatibility and confidence: Non-ASCII Windows paths, multi-input ONNX models, TensorRT validation, and distillation checkpoint handling receive targeted fixes that reduce failures in real-world workflows.

• Cleaner integrations: Platform and other log consumers can process progress updates directly instead of guessing which log lines represent progress bars.

• Release status: The package version is updated to 8.4.129.

What's Changed

• Add imread_unicode for non-ASCII paths to prevent downstream imread overriding for Windows by @Laughing-q in #21070
• Add validation in TensorRT test matrix by @Laughing-q in #22730
• Use ONNXBackend in ProfileModels for CPU speed benchmark by @Laughing-q in #23924
• Fully delegate end2end LiteRT exports to GPU accelerators by @Y-T-G in #25914
• Skip disk cleanup for Jetson builds by @glenn-jocher in #25930
• Parallelize batched image decoding in LoadImagesAndVideos by @JESUSROYETH in #25935
• Read exported head metadata only when the model is a path by @raimbekovm in #25929
• Keep export-format branches out of the head by @Y-T-G in #25924
• Add BF16 mixed-precision training support by @artest08 in #25931
• Report progress bar frames as state, not log lines by @Y-T-G in #25905
• Correct training guide claims and cross-link the Train & fine-tune group by @raimbekovm in #25927
• Parallelize FastSAM CLIP preprocessing by @JESUSROYETH in #25938
• Correct code-guide claims and verify every example against the package by @raimbekovm in #25934
• Delegate multi-dataset tuning to MultiTrainer by @glenn-jocher in #25937

Full Changelog: v8.4.128...v8.4.129

🌟 Summary

v8.4.128 improves OpenVINO reliability and batch inference, reduces RAM use during training, strengthens export behavior, and clarifies dataset, augmentation, and tracking workflows. 🚀

📊 Key Changes

• Synchronous OpenVINO batch inference — priority update ⚡
  OpenVINO now submits each input batch as a single synchronous request and consistently uses the LATENCY performance hint. Throughput implementations remain available internally, but mode selection is forced to latency-oriented execution to avoid hangs in AsyncInferQueue, particularly for dynamic INT8 batches on CPU systems.
• Improved OpenVINO efficiency and stability 🧠
  Benchmarks showed that one batched synchronous request was about twice as fast and used roughly one-third the RAM compared with splitting the batch into separate asynchronous requests. This should make OpenVINO exports more dependable in CI and production workloads, though throughput-focused applications may see different performance characteristics.
• TensorRT INT8 optimizations 🔧
  TensorRT 7–10 now keeps only the detection-head Sigmoid layers in higher precision instead of affecting every matching activation. TensorRT 11 no longer applies the unnecessary Sigmoid exclusion. This reduces model size and improves inference speed while preserving confidence calibration.
• More compatible ONNX exports 📦
  ONNX export now caps the opset at 18 to avoid CUDA execution falling back to CPU for unsupported operations. This prevents extra host-memory copies and should improve GPU inference consistency, especially for models such as RT-DETR.
• Safer RKNN exports 📱
  RKNN INT8 export now clearly rejects unsupported non-detection tasks and recommends FP16 instead. Documentation also adds updated YOLO26 FP16 and INT8 benchmarks for Rockchip devices.
• Lower RAM usage for cache='ram' 💾
  Cached images are stored in a shared contiguous memory buffer, preventing DataLoader workers from duplicating the cache during forked training. This should keep memory usage flatter when using multiple workers.
• SAM and FastSAM improvements 🎯
  • SAM auto-mask generation now encodes each crop once and reuses its features across point batches, reducing repeated computation.
  • FastSAM box and point prompts are clipped to image boundaries, preventing negative or out-of-range coordinates from silently selecting incorrect masks.
• More robust training and data utilities 🛠️
  • DDP now ignores externally set RANK and LOCAL_RANK values unless a real multi-process environment is detected.
  • Ray tuning correctly aggregates metrics across multiple datasets and reports the completed epoch.
  • Scalar indexing now preserves bounding-box formats and instance dimensions.
  • Ground-truth candidate selection is more consistent for very small boxes.
• Improved dataset workflows 📚
  NDJSON conversion now supports local image paths, while COCO JSON training documentation adds clearer requirements, cache warnings, and validation guidance.
• Documentation and usability updates ✍️
  Documentation now describes tracking as a mode that runs on detection, segmentation, pose, or OBB models; adds local Docker build instructions; expands task-specific K-Fold guidance; clarifies augmentation behavior; and documents Android ExecuTorch setup.
• AMP check weights are cached globally 📥
  The YOLO26n weights used only for AMP compatibility checks are now stored in the user configuration directory and reused across projects instead of being downloaded into the current working directory.

🎯 Purpose & Impact

• More reliable deployment: Synchronous OpenVINO execution reduces the risk of indefinite hangs and makes batch inference more predictable across Intel and AMD CPU environments. ✅
• Better performance on supported hardware: TensorRT and ONNX changes reduce unnecessary precision constraints and CPU fallback operations, potentially improving speed, memory usage, and GPU utilization.
• Lower training memory requirements: Shared RAM caching is especially beneficial for large datasets and multi-worker training.
• Fewer silent prediction errors: Prompt clipping and format-preserving indexing ensure that invalid coordinates or scalar selections do not produce misleading masks or improperly interpreted boxes.
• Clearer user guidance: Updated documentation makes it easier to choose the right task, prepare datasets, configure augmentations, build Docker images, and deploy models on mobile and edge hardware.
• Compatibility note: OpenVINO now prioritizes latency-safe synchronous execution over automatic throughput mode selection. Users seeking maximum throughput should benchmark their specific hardware and workload after upgrading.

What's Changed

• Refactor Docker quickstart docs by @ambitious-octopus in #18658
• Fix RKNN INT8 "no detections" and update benchmarks by @lakshanthad in #24703
• Add task-specific k-fold split guidance by @AffanBinFaisal in #24527
• Apply class-aware NMS in shared C++ postprocessing by @acgist in #23187
• Cache AMP check weights in global config directory by @gabe-zhang in #23664
• Handle environments where RANK and LOCAL_RANK variables are set outside of DDP context by @Y-T-G in #22724
• Fix cache='ram' memory leak via shared image buffer by @raimbekovm in #24673
• Support local image paths in NDJSON dataset conversion by @Y-T-G in #24357
• Add https://youtu.be/eotJNQwboWA to docs by @RizwanMunawar in #25910
• Fix Ray tuning across multiple datasets by @glenn-jocher in #25922
• Preserve bounding box format when indexing by @aswanth-07 in #25906
• Make GT side clamp in select_candidates_in_gts monotonic by @ShuaiLYU in #25912
• Scope the TensorRT INT8 Sigmoid exclusion to the detection head on both quantization paths by @onuralpszr in #25920
• Cap ONNX opset at 18 for ONNX Runtime CUDA kernel coverage by @Y-T-G in #25915
• Preserve instance axes during scalar indexing by @aswanth-07 in #25907
• Remove unreachable YOLODataset task assertion by @roel-klein in #20870
• Clip FastSAM box and point prompts to image bounds by @JESUSROYETH in #25916
• Correct InfiniteDataLoader iteration documentation by @PhuTd03 in #23578
• Remove unreachable resume-training branch by @noorchauhan in #24163
• Document ExecuTorch Android runtime setup by @blm25 in #22849
• perf(sam): encode each generate() crop once and reuse it across point batches by @JESUSROYETH in #25909
• Correct augmentation parameters and COCO JSON training example in data guides by @raimbekovm in #25919
• Describe tracking as a mode rather than a task in docs and READMEs by @raimbekovm in #25908
• Use synchronous OpenVINO batch inference by @glenn-jocher in #25921

New Contributors

• @roel-klein made their first contribution in #20870
• @acgist made their first contribution in #23187
• @blm25 made their first contribution in #22849
• @noorchauhan made their first contribution in #24163
• @PhuTd03 made their first contribution in #23578
• @aswanth-07 made their first contribution in #25906
• @gabe-zhang made their first contribution in #23664

Full Changelog: v8.4.127...v8.4.128

🌟 Summary

🚀 v8.4.127 makes exported YOLO models reliably load with the correct task and model family, while improving deployment stability, training recovery, and dataset documentation.

📊 Key Changes

• Correct task detection for exported models across all 20 formats by @artest08

  • Model loading now reads the task and architecture information embedded in export metadata instead of guessing from the filename or directory path.
  • Fixes segmentation exports being loaded as detection models, which previously caused masks to disappear.
  • Fixes pose exports losing keypoint outputs when weights are moved outside their training directory.
  • Correctly routes RT-DETR exports to the appropriate predictor, avoiding incorrect interpretation of its output format.
  • Applies consistently through both the Python API and CLI, making exported models safer to copy, deploy, and rename. 🎯

• Improved OpenVINO inference reliability

  • Uses the appropriate throughput setting for single-device dynamic batching.
  • Adds a targeted fix for INT8 dynamic-shape segmentation faults on Intel AMX CPUs.
  • Reduces the risk of crashes in affected CPU inference and CI environments.

• Safer training resume behavior

  • Resuming a run no longer replaces checkpoint weights with a custom pretrained model.
  • Preserves the actual model state, optimizer, scaler, EMA, and epoch from last.pt.
  • Restores distillation model handling during resume. 🔄

• More robust result handling

  • Semantic segmentation and depth result indexing now preserves complete dense maps instead of accidentally returning only one image row.
  • Improves iteration and indexing behavior for depth and semantic results.

• Expanded CoreML export support

  • nms=True now works for segmentation and pose exports in addition to detection.
  • Keeps masks and keypoints aligned with the boxes retained after suppression. 🍎

• Tracking and YOLOE fixes

  • Corrects DeepOCSORT OCR appearance matching so zero-overlap objects are not incorrectly matched.
  • Aligns TrackTrack new-track prediction and confirmation behavior with its reference implementation.
  • Fixes YOLOE prompt-free segmentation training crashes during final validation.

• Checkpoint and training reproducibility improvements

  • Custom Albumentations transforms are serialized as portable text representations rather than live Python objects.
  • YOLO26 documentation now explains the two-stage Objects365 pretraining and COCO fine-tuning process, embedded training logs, and checkpoint code revisions. 📚

• More direct Ultralytics Platform dataset access

  • Documentation now links official dataset pages for all 52 publicly hosted datasets, including detection, segmentation, pose, depth, and OBB resources.
  • Users can preview samples, inspect statistics, clone datasets, annotate, train, and deploy through the Ultralytics Platform.

• Documentation media delivery improvements

  • Documentation images and animations now use the CMS assets CDN, improving link previews and reliability across tools such as Slack and Telegram.
  • Several demonstrations were upgraded from static images to videos.

🎯 Purpose & Impact

• More dependable deployment: Exported models now retain their intended behavior even when renamed or moved into a different folder—a common step in production workflows.
• Correct predictions for specialized tasks: Segmentation masks, pose keypoints, and RT-DETR outputs are now handled by the correct predictors automatically.
• Fewer interrupted training runs: Resume behavior is safer, and OpenVINO fixes reduce crashes on supported Intel hardware.
• Better portability: Checkpoints containing custom augmentations can be loaded and shared more safely, although the original transform objects may still be needed when resuming with those augmentations.
• Broader Apple deployment options: CoreML users can now export segmentation and pose models with integrated NMS processing.
• Easier experimentation: Clearer YOLO26 training documentation and checkpoint metadata make it simpler to understand, reproduce, and fine-tune official models.
• A smoother dataset workflow: Exact Platform links make it faster to move from dataset discovery to annotation, training, and deployment. 🚀

What's Changed

• Link all official Platform datasets from Docs by @glenn-jocher in #25892
• Use correct OpenVINO throughput hint by @glenn-jocher in #25893
• Serve the docs media from the CMS assets CDN by @raimbekovm in #25884
• Ignore custom pretrained weights when resuming training by @Ajaysingh-2003 in #25816
• Preserve dense maps when indexing one-result Results by @JESUSROYETH in #25880
• fix: pass last-observation IoU to DeepOCSORT OCR appearance fusion by @JESUSROYETH in #25875
• Align TrackTrack New-track prediction and confirmation with the reference by @JESUSROYETH in #25877
• Fix YOLOE prompt-free segmentation training crash and re-enable the CLIP tests on Python 3.12 by @raimbekovm in #25867
• Document YOLO26 Objects365 pretraining and training curves by @Y-T-G in #25881
• Pick the YOLOE model and validator by task in the shared trainer by @raimbekovm in #25871
• Rewrite the YOLOE docs page and label the evaluation protocol on both LVIS result pages by @raimbekovm in #25856
• Fix OpenVINO INT8 dynamic-shape segfault on Intel AMX CPUs and replace dead docs link by @glenn-jocher in #25898
• Support nms=True for CoreML Segment and Pose exports by @SergioAlmeida29 in #25873
• Serialize custom augmentations losslessly in checkpoints by @Rahulbiradar9 in #25891
• Accept Path filenames in imread by @Kagura-Ahad in #25888
• Fit TQDM output to terminal width by @Y-T-G in #25166
• Stream live progress bar updates to console log consumers by @Y-T-G in #25900
• Load exported YOLO models with the correct task across all 20 formats by @artest08 in #25886

New Contributors

• @Ajaysingh-2003 made their first contribution in #25816
• @Kagura-Ahad made their first contribution in #25888

Full Changelog: v8.4.126...v8.4.127

🌟 Summary

v8.4.126 makes restricted checkpoint loading safer and up to 40% faster in concurrent environments, while simplifying RLE loss calculations and preserving backward compatibility.

📊 Key Changes

• 🔒 Thread-safe restricted checkpoint loading

  • Fixed a race condition when multiple threads loaded PyTorch checkpoints with restricted loading enabled.
  • The shared allow-list is now registered for the process lifetime instead of being temporarily removed when one thread finishes.
  • Added a regression test covering 32 concurrent restricted loads.

• ⚡ Faster restricted model loading

  • Only checkpoint globals actually referenced by a file are registered, rather than rebuilding a large allow-list for every load.
  • This reduces restricted-loading overhead by approximately 40% in affected cases.

• 🛡️ Improved compatibility with secure loading

  • Restricted loading now requires PyTorch functionality available from version 2.6 onward; older versions continue to fall back to standard loading behavior.
  • Normal, unrestricted loading remains unchanged.
  • Official YOLO26 checkpoints, including YOLO26x, are protected from failures caused by concurrent loading.

• 🧠 Simplified RLE prior calculation

  • Replaced the runtime multivariate distribution object with a simpler closed-form implementation.
  • Retained existing checkpoint buffers so models saved before v8.4.126 can still resume correctly.
  • Improved numerical behavior under mixed-precision training.

• 🏷️ Version update

  • Updated the Ultralytics package version from 8.4.125 to 8.4.126.

🎯 Purpose & Impact

• ✅ More reliable production inference: Multi-threaded services and Platform CPU workers can load models concurrently without intermittent checkpoint errors.
• 🚀 Reduced startup and loading time: Restricted checkpoint loading performs less unnecessary work, benefiting applications that frequently load models.
• 🔐 Maintained security benefits: Safe loading continues to limit checkpoint reconstruction to known, approved classes.
• 🔄 Backward compatible: Existing unrestricted workflows and older saved model checkpoints continue to work as before.
• 📉 Cleaner internal implementation: The RLE change removes unnecessary distribution-object overhead without changing the intended loss behavior.

What's Changed

• Make restricted checkpoint loading thread-safe and 40% faster, simplify RLE prior by @pderrenger in #25885

Full Changelog: v8.4.125...v8.4.126

🌟 Summary

Ultralytics 8.4.125 makes YOLO26 model startup significantly faster—up to 46% faster in fresh Python processes—while improving dependency loading and documentation consistency. 🚀

📊 Key Changes

• ⚡ Faster initial model loading (PR #25883 by @pderrenger):
  • TorchVision transforms are now imported only when a checkpoint actually contains them.
  • Matplotlib is loaded only when semantic-training label plots are requested.
  • Restricted checkpoint loading is substantially faster while retaining compatibility with older PyTorch versions.
  • Official YOLO26n and YOLO26x checkpoints showed approximately 44–46% faster model loading on an Apple M5 Pro.
  • Restricted loading improved from 451 ms to 94 ms for YOLO26n and from 517 ms to 162 ms for YOLO26x.
  • The common from ultralytics import YOLO import also became faster, improving from 558 ms to 471 ms.
• 🧪 Expanded loading validation:
  • Verified restricted loading across YOLO26 detection, classification, segmentation, pose, OBB, depth, and semantic checkpoints.
  • Confirmed multiple model types can be loaded sequentially in one process.
  • Tested compatibility with internal torch.nn checkpoint classes.
• 📦 Package version updated from 8.4.124 to 8.4.125.
• 🔗 Canonical Ultralytics URLs (PR #25879 by @pderrenger):
  • Removed trailing slashes from Ultralytics URLs across 136 files, including documentation, examples, source code, configuration, and workflow files.
  • Updated several outdated or redirected external links, including dataset, CUDA/cuDNN, MediaPipe, and PASCAL VOC references.

🎯 Purpose & Impact

• 🚀 Faster startup for developers and applications: Model-based scripts, services, notebooks, and command-line workflows begin running sooner, especially in fresh Python processes.
• 🪶 Lower unnecessary import overhead: Users who do not use semantic label plotting or serialized TorchVision transforms avoid loading those large optional dependencies.
• 🛡️ Safer and more reliable checkpoint handling: The updated loading logic preserves restricted loading behavior while supporting a broad range of official YOLO26 tasks and checkpoint formats.
• 🔄 Minimal migration impact: No major model architecture or training behavior changes are introduced; existing YOLO26 workflows should continue to work normally.
• 🧭 Cleaner documentation navigation: Consistent URL formatting and refreshed external links should reduce redirects, improve link maintenance, and provide more reliable access to guides.

What's Changed

• Remove trailing slashes from Ultralytics URLs by @pderrenger in #25879
• Speed up initial model loading by @pderrenger in #25883

Full Changelog: v8.4.124...v8.4.125

🌟 Summary

🚀 Ultralytics v8.4.124 restores reliable dynamic-size inference for exports with embedded NMS, while improving training stability, deployment compatibility, performance, and documentation.

📊 Key Changes

• Dynamic NMS exports restored — PR #25874

  • ONNX, OpenVINO, and TensorRT exports with dynamic=True and nms=True once again support runtime image heights and widths.
  • Exported NMS now consistently respects the configured max_det limit instead of using the number of anchors from the export image size.
  • NMS coordinate normalization now uses the actual input dimensions at inference time.
  • Dynamic ONNX OBB exports are padded appropriately so candidate selection is not limited by the traced image size.
  • The existing export test matrix remains unchanged, preserving broad backend coverage.

• Improved CoreML attention export

  • CoreML mlprogram exports now use a more compatible attention implementation to avoid GPU compilation crashes on recent Apple systems.
  • This is scoped to CoreML export and does not alter normal model execution.

• More efficient RT-DETR training and TensorRT inference

  • RT-DETR denoising queries are capped at the configured query budget, preventing dense images from causing excessive memory use.
  • RT-DETR TensorRT Top-K processing now uses grouped selection for faster candidate filtering.

• Training reliability fixes

  • Resumed training now preserves the weights from the checkpoint instead of accidentally reloading the original pretrained weights.
  • Training seeds now reach dataloader workers, making different seeds produce different augmentation and sampling sequences while preserving reproducibility.
  • Dataset construction no longer mutates shared training configuration values.
  • Deterministic training settings are always cleared when training finishes or fails.
  • AutoBatch now raises a clear error when no tested batch size fits, rather than silently falling back to an unrelated default.

• Prediction and result-processing improvements

  • Reusing a model for predict() or track() no longer carries filters such as classes, max_det, or NMS settings into later calls.
  • GPU-to-CPU transfers in Results.save_txt(), save_crop(), and summary() are consolidated, reducing per-object synchronization overhead.
  • Segmentation mask encoding now transfers data to the CPU more efficiently.
  • Depth tensor inputs now use the expected BGR image order.
  • Coordinate restoration now handles stretched, nonuniform resizing correctly.

• Export and platform updates

  • ONNX INT8 export uses less peak memory by releasing intermediate graphs earlier.
  • TensorRT dynamic optimization profiles now handle dynamic input dimensions more safely.
  • macOS dependency constraints avoid affected NumPy releases associated with Accelerate warnings.
  • Platform documentation now describes depth datasets, depth-map viewing, training requirements, and updated workflow behavior.
  • Hailo documentation now reflects current hardware names, AI HAT+ support, compiler generations, precision notes, and fixed-shape deployment requirements.
  • Tracking documentation now identifies TrackTrack as the default tracker for documented workflows.
  • CoreML documentation includes an updated YOLO26 INT8 deployment tutorial.

🎯 Purpose & Impact

• More dependable deployment across input sizes 📐 — Dynamic exported models can process images at runtime sizes without silently bypassing confidence filtering or the max_det limit.
• Correcter detections and coordinates 🎯 — Runtime-aware normalization and nonuniform scaling fixes help ensure boxes, keypoints, and OBB predictions remain accurately positioned.
• Lower memory use and faster execution ⚡ — RT-DETR query capping, grouped Top-K selection, consolidated GPU transfers, and improved ONNX memory handling benefit both training and inference.
• Safer long-running workflows 🔁 — Reused prediction models, resumed training, seeded dataloaders, and deterministic settings now behave more predictably.
• Better edge and Apple deployment 🍎📦 — CoreML and Hailo export guidance and compatibility improvements make it easier to move YOLO models to supported hardware.
• Clearer user guidance 📚 — Updated documentation better reflects actual defaults, supported tasks, deployment limitations, and Platform workflows.

What's Changed

• Add https://youtu.be/KcTSdIUYcVE to docs by @RizwanMunawar in #25844
• Document the effective export defaults for verbose and imgsz by @raimbekovm in #25831
• Fix depth tensor source color order by @ahmet-f-gumustas in #25833
• Cap RT-DETR denoising queries at the query budget by @artest08 in #25834
• Document the defaults Solutions actually apply by @raimbekovm in #25835
• Align Platform workflow diagrams with the product by @raimbekovm in #25840
• Use perf_counter() for ProfileModels and time_sync() latency measurement by @raimbekovm in #25845
• Document Windows multi-GPU DDP limitation by @diaz3z in #25850
• Fix nonuniform ratio-pad coordinate scaling by @tandede in #25855
• Avoid per-box GPU syncs in Results.save_txt/save_crop/summary by @JESUSROYETH in #25842
• Fix hyp aliasing so dataset construction cannot mutate the shared cfg by @SergioAlmeida29 in #25863
• fix(engine): keep resumed checkpoint weights when resuming training by @Sigwendice in #25854
• Fix macOS metadata directory cleanup by @uczltw6 in #25857
• Always unset deterministic state when training ends by @Y-T-G in #25862
• perf: pack mask RLE indices and transfer to CPU once in SegmentationValidator by @JESUSROYETH in #25853
• Fix nms=True export silently dropping detection filtering under dynamic=True by @JESUSROYETH in #25843
• Document depth dataset support across the Platform docs by @laodouya in #25849
• Restore the copy_paste candidate fraction by @raimbekovm in #25810
• Exclude affected NumPy versions on macOS by @raimbekovm in #25848
• Fix CoreML GPU compile crash in attention export by @Y-T-G in #25860
• Reduce peak memory during ONNX INT8 export by @amanharshx in #25817
• Map semantic mask labels with a 256-entry lookup table by @raimbekovm in #25813
• Speed up RT-DETR top-k on TensorRT by @artest08 in #25755
• Raise a clear error in autobatch() when no candidate batch size fits by @JESUSROYETH in #25739
• Fix #25764: Don't send local project paths as Platform slugs by @altf4-games in #25824
• Fix predict()/track() kwargs sticking across reused calls on the same model by @JESUSROYETH in #25731
• Refresh Hailo integration docs: hardware overview, product naming, and resource links by @eldadr in #25786
• Make seed reach dataloader workers so augmentations vary between runs by @yentur in #25815
• Add optional channel_divisor model YAML key for channel rounding control by @HussainNizamani in #25423
• Add https://youtu.be/3Z0_Fxhm030 to docs by @RizwanMunawar in #25866
• Simplify Platform project path slugging by @glenn-jocher in #25869
• Restore balanced Hailo deployment guidance by @glenn-jocher in #25868
• Revert optional channel_divisor model YAML key by @glenn-jocher in #25870
• Restore dynamic image sizes for NMS exports by @glenn-jocher in #25874

New Contributors

• @HussainNizamani made their first contribution in #25423
• @SergioAlmeida29 made their first contribution in #25863
• @yentur made their first contribution in #25815
• @eldadr made their first contribution in #25786
• @tandede made their first contribution in #25855
• @uczltw6 made their first contribution in #25857
• @altf4-games made their first contribution in #25824
• @Sigwendice made their first contribution in #25854

Full Changelog: v8.4.123...v8.4.124