ultralytics

🌟 Summary

Ultralytics v8.4.60 is mainly about adding ONNX INT8 export 🎉, making it easier to create smaller, faster deployment models with built-in calibration support, while also including a few helpful export, training, and documentation fixes.

📊 Key Changes

• 🚀 Major new feature: ONNX int8=True export

  • You can now export models like YOLO26 to INT8 ONNX using ONNX Runtime static quantization.
  • This uses the same familiar export flow as other INT8 formats, including data for calibration dataset selection and fraction for using only part of the dataset.
  • Exported files now clearly save as *_int8.onnx.

• 🔄 Shared INT8 calibration pipeline

  • The new ONNX INT8 export reuses the existing calibration dataloader logic instead of introducing a separate workflow.
  • Calibration reader code is now shared between ONNX and QNN export, reducing duplication and making behavior more consistent.

• 📘 Much better ONNX export documentation

  • ONNX INT8 support was added across export docs, ONNX integration docs, exporter examples, macros, reference pages, and the tutorial notebook.
  • This makes the new feature easier to discover and use correctly.

• ⚙️ RKNN export now supports the standard half argument

  • RKNN exports now officially support half=True, and this becomes the default floating-point path for supported Rockchip hardware.
  • This makes RKNN behavior more consistent with other export formats.

• 🐛 Segmentation training fix for polygons on image borders

  • A fix to segment2box ensures polygon points lying exactly on image edges are no longer dropped.
  • This helps prevent bad bounding boxes and training issues for objects that touch the image border.

• 📝 Auto-annotate docs updated

  • Documentation now correctly includes SAM 3 in the supported SAM model list.
  • The documented default output_dir for auto-annotation was corrected.

• 🧹 Docs metadata cleanup

  • Placeholder documentation metadata was replaced with meaningful defaults, improving generated docs quality.

🎯 Purpose & Impact

• 🎯 Faster and lighter ONNX deployment

  • The headline feature is ONNX INT8 export, which can help reduce model size and improve inference efficiency on supported runtimes and hardware.
  • This is especially useful for edge devices, production services, and resource-constrained deployments.

• 🛠️ Simpler export workflow

  • Users already familiar with Ultralytics INT8 export options will find ONNX INT8 works in a similar way, so there is less new syntax or tooling to learn.

• 🔒 More reliable maintenance and consistency

  • Sharing calibration logic between ONNX and QNN reduces duplicate code, which usually means fewer bugs and easier long-term support.

• 📈 Better user experience for deployment

  • Clearer docs and examples should make it easier for both new and experienced users to adopt ONNX INT8 export successfully.

• 🤖 Improved hardware export support

  • The RKNN half=True update helps Rockchip deployments behave more predictably and aligns them better with common export expectations.

• 🖼️ More accurate segmentation training

  • The border-polygon fix can improve training data handling for segmentation datasets where objects touch image edges, avoiding accidental quality loss.

Overall, v8.4.60 is a deployment-focused release 🌍, with ONNX INT8 export as the standout improvement and several supporting fixes that improve reliability, documentation, and hardware export consistency.

What's Changed

• Allow half arg for RKNN export by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24660
• Fix SAM model list and default output dir in auto-annotate docs by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/24662
• Fix segment2box dropping polygon vertices on image border by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/24655
• ultralytics 8.4.60 ONNX INT8 export by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24666

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.59...v8.4.60

🌟 Summary

Ultralytics v8.4.59 improves Rockchip RKNN export with safer and clearer INT8 quantization support, making deployment on Rockchip edge devices more reliable and easier to use 🚀

📊 Key Changes

• ✅ RKNN INT8 calibration file is now validated before export proceeds

  • The release’s main follow-up change from @glenn-jocher checks that the internally generated calibration image list actually exists before loading the RKNN Toolkit.
  • This helps catch setup issues earlier, with clearer error messages instead of failing later in a more confusing way.

• 🧠 Simpler RKNN INT8 workflow for users

  • Users continue to provide a standard YOLO dataset YAML with data=... during export.
  • Ultralytics now handles conversion of that dataset into RKNN’s internal dataset.txt calibration file automatically.
  • dataset.txt stays an internal implementation detail, not something users are expected to manage directly.

• ⚡ New RKNN INT8 export support introduced in this release cycle

  • RKNN export now supports:
    • int8=True
    • data=...
    • fraction=...
  • This enables INT8-quantized RKNN model export, which is especially important for Rockchip devices that require INT8 models.

• 📱 Expanded support for INT8-only Rockchip chips

  • Devices such as RV1103, RV1106, RV1103B, and RV1106B are now supported through INT8 export.
  • The exporter also warns and auto-adjusts when a selected chip requires int8=True.

• 📚 Documentation and export tables updated

  • RKNN docs now clearly explain:
    • floating-point export behavior for FP16-capable targets
    • when int8=True is required
    • use of data and fraction for calibration
  • Export reference tables and tutorial content were updated to reflect the new RKNN options.

• 🛡️ Better RKNN API error handling

  • Internal RKNN API calls now check return values more carefully.
  • Failures in config, ONNX loading, build, or export are surfaced as clearer runtime errors.

🎯 Purpose & Impact

• More reliable RKNN exports 🔒

  • Validating the calibration file up front reduces frustrating export failures and makes debugging easier.

• Easier INT8 quantization for everyone 🙌

  • Users can stick to the familiar Ultralytics pattern of passing data=coco8.yaml or another dataset YAML, without learning RKNN-specific file conventions.

• Better support for edge deployment on Rockchip hardware 📦

  • INT8 export opens the door to more Rockchip NPUs, including chips that cannot use floating-point RKNN models.

• Cleaner user experience 🧹

  • The release hides low-level RKNN details from users and keeps the interface aligned with other Ultralytics export formats.

• Potential performance and compatibility gains 🚀

  • INT8 models are often smaller and better suited for constrained edge devices.
  • For supported Rockchip targets, this can improve deployability and help users target hardware that previously wasn’t practical with RKNN export.

• Lower risk of misconfiguration ⚠️

  • Clearer warnings, updated docs, and improved validation make it less likely that users will choose unsupported settings by accident.

If helpful, I can also provide a 1-paragraph release note version or a developer-focused changelog summary for v8.4.59.

What's Changed

• ultralytics 8.4.59 Rockchip RKNN INT8 export option by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24656
• Validate RKNN INT8 calibration dataset path by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24659

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.58...v8.4.59

🌟 Summary

Ultralytics v8.4.58 is mainly a reliability-focused release 🛠️ that improves how export formats are tested in CI, making model export support more stable, isolated, and easier to maintain—especially for specialized formats.

📊 Key Changes

• 🚦 Export testing was redesigned around isolated environments

  • The biggest update in this release, from PR #24649 by @glenn-jocher, separates export tests into dedicated environments instead of testing everything in one shared CI setup.
  • This is especially important for formats with conflicting dependencies, such as TensorFlow, CoreML, MNN, NCNN, ExecuTorch, IMX, RKNN, Axelera, and DEEPX.

• 🧩 Each export format now declares its own environment

  • The internal export_formats() registry now includes environment metadata for every export target.
  • This lets the test system know exactly where each export should run.

• 🧪 New --export-env test selection

  • CI no longer relies on manually skipping individual export tests.
  • Instead, tests are selected by export environment, which is cleaner and much less error-prone.

• ⚙️ New automated isolated export job

  • A new IsolatedExports CI workflow now builds dedicated virtual environments and runs smoke tests plus export tests inside them.
  • This replaces the older shell-script-based slow test coordination.

• 📦 Export dependencies were split into smaller groups

  • The release breaks the old monolithic export dependencies into targeted extras like:
    • export-base
    • export-tensorflow
    • export-coreml
    • export-executorch
    • export-deepx
    • export-legacy-torch
  • This keeps base installs leaner and avoids pulling in heavy packages unless they are truly needed.

• ✅ Added checks to prevent export registry mistakes

  • New tests ensure:
    • every export environment has a smoke test
    • every export format points to a valid registered environment

• 🤖 RKNN export test added

  • A dedicated test for RKNN export was added, improving coverage for Rockchip deployment workflows.

• 🐳 Docker cleanup improved

  • Export Docker images now clean caches more aggressively, helping reduce leftover build artifacts and keeping images/workflows tidier.

🎯 Purpose & Impact

• 🔒 More reliable exports

  • Users who export YOLO models to many different formats should see fewer CI-related regressions slipping through.
  • In simple terms: export support is being tested in a way that better matches real-world usage.

• 🚀 Faster and leaner development workflows

  • Base CI jobs now install a smaller export stack, which should reduce unnecessary dependency load and make routine testing more efficient.

• 🧹 Simpler maintenance for the Ultralytics team

  • The new registry-driven setup replaces fragile manual skip lists and shell workarounds.
  • That means adding or maintaining export formats should be easier going forward.

• 🌍 Better support for specialized deployment targets

  • Formats used for edge and hardware-specific deployment—like IMX, RKNN, Axelera, DEEPX, and ExecuTorch—now get clearer, dedicated validation paths.

• 📌 No major new model architecture in this release

  • This release does not introduce a new YOLO model or major end-user training feature.
  • Its main impact is improved stability, compatibility, and confidence in exporting existing models for deployment.

• 🧠 Good news for production users

  • If you depend on exporting YOLO models to mobile, embedded, browser, or accelerator-specific formats, this release reduces the chance that one format breaks another behind the scenes.

If you'd like, I can also turn this into a shorter changelog-style summary or a more technical developer-focused breakdown.

What's Changed

• ultralytics 8.4.58 Isolated CI export environments by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24649

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.57...v8.4.58

🌟 Summary

🚀 Ultralytics v8.4.57 is mainly a Python 3.13 readiness release, making CI and export workflows more robust across platforms, while also adding a meaningful RT-DETR CoreML export upgrade and a semantic segmentation mask alignment fix.

📊 Key Changes

• 🐍 Python 3.13 CI is now the main focus of this release

  • CI testing moved to Python 3.13 across Linux, Windows, macOS, ARM, GPU, and benchmark jobs.
  • Added official Python 3.13 package support metadata.
  • Updated dependency rules for export-related packages like TensorFlow, ONNX, CoreML, NumPy, and protobuf.

• 🛡️ Export and slow-test CI became much more resilient

  • Added a dedicated slow-test runner script with better diagnostics.
  • CI now captures clearer logs for crashes, memory/resource issues, and dependency conflicts instead of failing silently.
  • Extra isolation was added for export tests to reduce flaky failures caused by parallel runs.

• 🚫 Python 3.13 compatibility guards were added for unsupported export stacks

  • TensorFlow export is blocked on macOS + Python 3.13 due to upstream wheel issues.
  • TensorFlow.js export is blocked on Python 3.13 because its current dependency stack is not yet compatible.
  • Axelera export is skipped on Python 3.13 for now.
  • DEEPX CI validation stays on Python 3.12 because its compiler wheels are not yet available for 3.13.

• 📦 RT-DETR CoreML export improved significantly

  • RT-DETR-L can now export to CoreML mlprogram more reliably.
  • Core parts of the decoder were adjusted to preserve precision, especially where FP16 could cause output drift.
  • New tests were added to validate RT-DETR CoreML export.

• 🎯 Semantic segmentation masks are now better aligned

  • Fixed a subtle issue where masks could appear slightly offset because of sub-pixel LetterBox padding.
  • Predictions are now resized more carefully before padding removal, leading to cleaner masks on the original image.

• 📚 Docs and platform updates

  • Added new Ultralytics Platform integrations docs for importing from Ultralytics HUB and Roboflow.
  • Large refresh of DEEPX documentation and branding consistency.
  • Fixed doc inaccuracies and broken links across Comet, TensorRT, Edge TPU, TF.js, ExecuTorch, OBB docs, and more.
  • Explorer legacy snippets were corrected to use YOLO11 models that actually match the pinned older package version.

🎯 Purpose & Impact

• ✅ Better future-proofing for users on new Python versions

  • If you’re moving to Python 3.13, this release makes Ultralytics much more prepared for that transition.
  • Users should see more predictable installation, testing, and export behavior.

• 🔍 Much easier debugging for maintainers and advanced users

  • CI failures should now be more informative instead of abruptly dying with little explanation.
  • This helps speed up fixes and makes release quality stronger over time.

• 📤 Safer export experience across many backends

  • Instead of letting unsupported combinations fail mysteriously, Ultralytics now blocks or skips them with clearer reasoning.
  • That means less wasted time for users trying unsupported exports.

• 🍎 RT-DETR deployment on Apple ecosystems gets more practical

  • CoreML export improvements make RT-DETR more usable for iPhone, iPad, and Apple device workflows.
  • This is one of the biggest model-related improvements in the release.

• 🖼️ Semantic segmentation results should look cleaner

  • Users working with segmentation may notice more accurate mask placement, especially around edges and padding-sensitive cases.

• 🌐 Easier migration into the Ultralytics Platform

  • New docs for importing from Ultralytics HUB and Roboflow make it simpler to bring existing datasets and assets into the Ultralytics Platform 📥

Overall, v8.4.57 is less about introducing a brand-new model and more about stability, compatibility, and deployment polish 🛠️✨—especially for Python 3.13, export workflows, and RT-DETR/CoreML users.

What's Changed

• Enable RT-DETR CoreML export and protect decoder ops from FP16 drift by @artest08 in https://github.com/ultralytics/ultralytics/pull/24510
• docs: Fixing text for DEEPX brand rules & guidlines by @deepx-dgkim in https://github.com/ultralytics/ultralytics/pull/24535
• Fix semantic segmentation mask offset from sub-pixel LetterBox padding by @lmycross in https://github.com/ultralytics/ultralytics/pull/24619
• Fix race condition in test_export by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/24609
• Fix 404 links in Comet and TensorRT docs by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24634
• Add DEEPX vs PyTorch benchmarks pic by @lakshanthad in https://github.com/ultralytics/ultralytics/pull/24636
• Correct OBB angle range note in docs by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/24534
• Improve readability of distance calculation documentation by @Sunny11092003 in https://github.com/ultralytics/ultralytics/pull/24625
• Fix classify CustomizedValidator example augmenting the val set by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/24532
• Use yolo11n in explorer legacy snippets to match 8.3.11 pin by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/24548
• Change CI OS from cpu-latest to ubuntu-latest by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24641
• Docs: Platform Integrations by @sergiuwaxmann in https://github.com/ultralytics/ultralytics/pull/24642
• ultralytics 8.4.57 Python 3.13 CI by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24640

New Contributors

• @Sunny11092003 made their first contribution in https://github.com/ultralytics/ultralytics/pull/24625
• @deepx-dgkim made their first contribution in https://github.com/ultralytics/ultralytics/pull/24535

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.56...v8.4.57

🌟 Summary

Ultralytics v8.4.56 improves QNN export reliability by fixing compatibility with newer onnxruntime-qnn packages, especially on Linux x86-64. 🚀

📊 Key Changes

• Fixed QNN export for built-in provider wheels 🔧
  Ultralytics now detects whether onnxruntime-qnn already includes QNNExecutionProvider internally, instead of always trying to register it as a separate plugin.

• Avoids export failures on some Linux setups 🐧
  This prevents a known failure on Linux x86-64 wheels where plugin registration could crash with an undefined symbol error.

• Supports both QNN packaging styles 📦
  The export flow now works with:

  • plugin-based QNN wheels, where the provider must be registered manually
  • built-in QNN wheels, where registration is not needed

• Smarter QNN session creation 🧠
  If the provider is built in, Ultralytics now launches the ONNX Runtime session directly with the QNN provider. If it is plugin-based, it keeps the previous registration path.

• Documentation/comments updated for clarity 📝
  Internal descriptions were updated to reflect that QNN may be exposed either as a plugin or as a built-in ONNX Runtime provider.

🎯 Purpose & Impact

• Makes QNN export more robust ✅
  Users exporting models to Qualcomm QNN should see fewer environment-specific errors and smoother exports.

• Improves compatibility with newer runtime wheels 🔄
  This is especially helpful as onnxruntime-qnn packaging evolves across platforms and releases.

• Reduces setup friction for deployment 📱
  If you are exporting YOLO models like YOLO26 for Qualcomm and edge hardware workflows, this change lowers the chance of export-time breakage.

• No major model architecture changes in this release 🤖
  v8.4.56 is mainly a stability and export compatibility update, not a new model or training feature release.

• Most relevant for advanced deployment users ⚡
  If you use standard training and inference, this release may not change much day to day. But for users targeting QNN / Qualcomm deployment pipelines, it should have a meaningful positive impact.

🙌 Bottom line

This release is a small but important deployment fix: it makes QNN export smarter and more portable by handling both old and new onnxruntime-qnn layouts correctly, with the biggest benefit for Linux-based export environments.

What's Changed

• Fix QNN export with built-in provider wheels by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24618

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.55...v8.4.56

🌟 Summary

Ultralytics v8.4.55 is a reliability-focused release that mainly fixes ONNX/QNN export environments so existing onnxruntime variants are no longer accidentally overwritten, while also improving deployment docs, especially for Hailo and YOLO26 task support 🚀

📊 Key Changes

• 🛠️ Major export fix for ONNX/QNN environments

  • The most important change in this release is a fix in ONNX export dependency checking by @glenn-jocher.
  • Ultralytics now properly accepts alternative onnxruntime package variants such as:
    • onnxruntime-qnn
    • onnxruntime-directml
    • onnxruntime-openvino
    • onnxruntime-training
  • Previously, export checks could mistakenly think ONNX Runtime was missing and auto-install standard onnxruntime, which could overwrite specialized builds and break export or inference.

• 📦 Safer QNN export workflows

  • This especially helps users exporting for Qualcomm QNN.
  • It prevents ABI/provider conflicts that could cause failures during QNN-related ONNX export steps.

• 📘 New Hailo deployment guide

  • Added a detailed new guide for deploying Ultralytics YOLO models to Hailo hardware through an ONNX-to-HEF workflow.
  • The docs explain that Hailo is not a direct model.export() target, but can still be used through the external Hailo toolchain.

• 🧠 Broader and clearer YOLO26 task documentation

  • Many docs were updated to better distinguish:
    • instance segmentation
    • semantic segmentation
  • YOLO26 support tables, CLI docs, platform docs, and integration pages now more clearly reflect that semantic segmentation is supported.

• ⚡ Faster and more stable CI testing

  • Test infrastructure now uses parallel pytest runs with pytest-xdist.
  • Shared test assets are cached in advance, and test outputs are isolated to avoid conflicts.
  • This should help improve release quality and reduce flaky test failures.

• 📈 Improved Comet documentation

  • Added missing COMET_* environment variable docs and clarified online/offline logging behavior.

• 🔄 Minor maintenance updates

  • GitHub Actions cache was updated from actions/cache@v4 to @v5.
  • Small wording and docstring cleanups were made across the project.

🎯 Purpose & Impact

• ✅ More reliable exports for advanced runtimes

  • If you use specialized ONNX Runtime builds, this release is important.
  • It reduces the risk of Ultralytics auto-installing the wrong package and breaking your setup.

• 📱 Better support for edge and hardware-specific deployment

  • QNN users should see fewer export-related environment issues.
  • Hailo users now have clearer guidance for getting YOLO models onto supported devices.

• 🧩 Less confusion around segmentation tasks

  • The docs now do a better job of explaining the difference between instance segmentation and semantic segmentation, which is especially helpful for newer users.

• 🛡️ Better overall stability

  • Faster and better-isolated test pipelines should help catch regressions earlier and improve release reliability for everyone.

• 🌍 Easier adoption of YOLO26 capabilities

  • The documentation now more consistently highlights that YOLO26 supports six task types, including semantic segmentation, making the platform easier to understand and use.

In short: v8.4.55 is mainly a stability and compatibility release 🔧—especially valuable for users exporting to ONNX/QNN or working with specialized runtime builds—while also making deployment and YOLO26 task support clearer for the broader community.

What's Changed

• Update semantic task support docs by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24597
• Enable parallel CI tests via pytest-xdist and fix export/train race conditions by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/24565
• Document missing COMET_* environment variables by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/24600
• Hailo export guide by @ambitious-octopus in https://github.com/ultralytics/ultralytics/pull/24596
• Bump actions/cache from v4 to v5 in /.github/workflows by @UltralyticsAssistant in https://github.com/ultralytics/ultralytics/pull/24604
• ultralytics 8.4.55 Accept onnxruntime variants in ONNX export requirements check by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24611

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.54...v8.4.55

🌟 Summary

Ultralytics v8.4.54 is led by a big new deployment feature: local Qualcomm QNN export for Snapdragon devices 📱⚡, plus several quality-of-life improvements for Solutions, export docs, tuning compatibility, and data-loading reliability.

📊 Key Changes

• 🚀 New qnn export format for Qualcomm Snapdragon devices

  • You can now export YOLO models to Qualcomm QNN with format="qnn".
  • This compiles a model locally into a QNN-ready package using ONNX Runtime’s QNN Execution Provider.
  • No Qualcomm account, no cloud upload, and no separate Qualcomm SDK download are required.
  • Works with Snapdragon targets including CPU, GPU, and Hexagon NPU paths.
  • Example:

    from ultralytics import YOLO
    
    model = YOLO("yolo26n.pt")
    model.export(format="qnn")
    

• 📦 New QNN backend support for inference/validation workflows

  • Exported QNN models can now fit into the normal Ultralytics workflow more cleanly.
  • QNN is registered as a supported export target, with dedicated backend and docs support.
  • End-to-end export is automatically disabled for QNN where unsupported, helping avoid invalid configurations.

• 📘 Major export documentation cleanup and standardization

  • Integration docs across many formats were reorganized into a clearer Export / Predict / Validate structure.
  • A full new guide for Qualcomm QNN was added.
  • This makes deployment paths easier to understand across formats like ONNX, TensorRT, OpenVINO, NCNN, TFLite, CoreML, and more.

• 🖼️ imgsz added to all Ultralytics Solutions

  • All Solutions now support the imgsz argument, giving users direct control over inference input size.
  • This also improves logging so reported input shapes better reflect the real model input.

• 🛠️ Better robustness for segmentation and image caching

  • Fixed semantic segmentation mask loading when masks appear as (H, W, 1) instead of plain single-channel arrays.
  • Fixed stale .npy image cache issues that could cause grayscale models to accidentally load RGB data.

• 🧠 Tracking utility optimization

  • Reduced some SciPy dependency usage in tracking internals by replacing parts with NumPy equivalents.
  • This helps simplify dependencies and improve maintainability.

• 🎯 Ray Tune compatibility improvement

  • Updated tuner trial ID handling to work correctly with Ray 2.7+ while keeping support for older versions.

• 🧾 Small but useful documentation updates

  • Validation docs now include show_labels and show_conf.
  • Prediction docs now explicitly list .jpeg2000 as a supported image format.

• 🔬 Semantic segmentation loss dtype fix

  • Improved dtype consistency in segmentation loss computation, especially helpful for mixed-precision training paths.

🎯 Purpose & Impact

• 📱 Big win for edge/mobile deployment

  • The new QNN export makes it much easier to deploy YOLO on Qualcomm Snapdragon hardware.
  • This is especially valuable for mobile, embedded, automotive, and low-power AI use cases.

• ⚡ Faster path to production on Snapdragon

  • Users can now go from a standard YOLO model to a Snapdragon-ready export without extra accounts or vendor tooling setup.
  • That lowers friction for developers evaluating on-device AI.

• 🧩 More consistent deployment experience

  • Standardized export docs make it easier for both beginners and experienced users to compare formats and follow the right steps.

• 🎛️ More control in packaged Solutions

  • Adding imgsz to Solutions gives users a simple way to trade off speed vs. accuracy, just like in normal YOLO inference.

• 🛡️ More reliable training and inference

  • The mask-loading and grayscale cache fixes reduce confusing failures caused by edge-case dataset formats or stale cache files.

• 🔄 Better compatibility with modern tooling

  • Ray Tune improvements help users keep hyperparameter tuning workflows working across newer environments.

• 👀 Cleaner debugging and validation visuals

  • The documented show_labels and show_conf options make validation outputs easier to customize for analysis.

Overall, v8.4.54 is a strong release for anyone interested in device deployment, especially on Qualcomm Snapdragon, while also tightening up reliability and usability across the broader Ultralytics ecosystem 🚀

What's Changed

• Docs: add show_labels and show_conf val args by @K-saif in https://github.com/ultralytics/ultralytics/pull/24571
• Scope scipy imports and optimize embedding_distance with NumPy by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/24572
• Revert _gaussian_smooth changes by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/24575
• Match aux_loss dtype to ce_loss in SemanticSegmentationLoss by @ahmet-f-gumustas in https://github.com/ultralytics/ultralytics/pull/24576
• Add imgsz support to all Ultralytics Solutions by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/24579
• Fix: Squeeze single-channel 3D mask arrays from patched cv2.imread by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/24582
• Fix grayscale test failure from stale .npy cache files by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/24583
• Add jpeg2000 to predict images formats table by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/24588
• Fix tuner trial ID retrieval for Ray 2.7+ compatibility by @lmycross in https://github.com/ultralytics/ultralytics/pull/24590
• ultralytics 8.4.54 Qualcomm QNN export by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24591

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.53...v8.4.54

🌟 Summary

Ultralytics v8.4.53 mainly improves training reliability on NVIDIA GPUs by automatically recovering from more CUDA memory-related failures, while also polishing semantic segmentation stability, documentation clarity, and CI robustness 🚀

📊 Key Changes

• Smarter GPU memory recovery during training 🧠⚡
  The biggest change in this release, from PR #24569 by @glenn-jocher, expands the existing auto-retry logic for large-batch GPU training.

  • Previously, Ultralytics could retry after a standard CUDA out-of-memory error.
  • Now it also retries for related backend memory failures such as:
    • CUDNN_STATUS_INTERNAL_ERROR
    • “unable to find an engine to execute this computation”
  • When this happens during the first epoch on a single GPU, training can automatically reduce batch size and continue instead of crashing.

• Semantic segmentation validation is more reliable 🖼️✅
  PR #24552 fixes issues in semantic validation so metrics are preserved correctly after evaluation.

  • Prevents stats from being cleared too early
  • Makes mIoU handling more robust when some classes are missing
  • Avoids NaN-style metric problems in sparse-class cases

• Segmentation loss is numerically more stable 🛠️📈
  PR #24554 improves Dice loss stability by forcing key intermediate calculations to use float32.

  • Helps avoid precision-related instability
  • Especially useful in mixed-precision or lower-precision training setups

• Much better docs for prediction outputs, especially semantic segmentation 📚
  PR #24558 is a major documentation improvement.

  • Clearly explains the Results object for each task
  • Documents the new/important semantic_mask output in detail
  • Adds comparisons between instance segmentation and semantic segmentation
  • Clarifies output shapes, dtypes, masks, polygons, boxes, and per-task behavior

• Broader YOLO26 and semantic task visibility across docs and metadata 🌍
  Several updates improve how Ultralytics presents current capabilities:

  • Package metadata now reflects YOLO26, Platform, and oriented object detection
  • Project classifier changed to Production/Stable
  • Many docs now consistently mention semantic segmentation as a first-class supported task

• CI and test workflow became more stable 🧪
  Multiple changes reduce avoidable test failures:

  • IMX export test is skipped in broader CI where it was causing instability
  • Training-related tests are skipped on Raspberry Pi, where they are not practical

• Small but useful documentation and packaging fixes ✨

  • README task banner now renders correctly on PyPI
  • OBB validation docs now note the true default confidence used to reduce memory usage
  • Minor doc fixes for tracking links, benchmark hardware, glossary links, and formatting

🎯 Purpose & Impact

• Fewer frustrating training crashes on GPUs 💪
  If you train with aggressive batch sizes, this release can help Ultralytics recover automatically from more real-world CUDA memory failures instead of stopping unexpectedly.

• Better experience for users pushing hardware limits 🚀
  Large-batch training on single GPUs should now be more forgiving, especially when failures come from CUDA backend libraries rather than a plain out-of-memory exception.

• More dependable semantic segmentation workflows 🧠
  Users working with semantic segmentation should see:

  • more stable validation metrics
  • fewer confusing missing/invalid values
  • improved training stability from safer Dice loss calculations

• Easier to understand model outputs 🔍
  The new docs make it much clearer what prediction results contain for each task, which is especially helpful for:

  • beginners trying to use Results
  • developers building pipelines around semantic outputs
  • teams comparing instance vs semantic segmentation behavior

• Better confidence in production use ✅
  The “Production/Stable” metadata and improved CI signal a continued push toward reliability for real deployments.

• Low risk release with practical quality-of-life improvements 🙂
  This is not a major architecture or model release, but it delivers meaningful gains in stability, clarity, and usability, especially for training and semantic segmentation users.

What's Changed

• Clean up semantic CI follow-up by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24544
• Fix tasks banner not rendering on PyPI by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/24545
• Update pyproject metadata for YOLO26 and Platform by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24547
• Fix Cityscapes glossary link by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24550
• Fix semantic docs typo by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/24551
• Fix semantic nt_per_image printing issue by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/24552
• Fix NaN Dice loss during validation by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/24554
• Skip training on Raspberry Pi Test CIs by @lakshanthad in https://github.com/ultralytics/ultralytics/pull/24556
• Add segmentation results output details to task docs by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24558
• Note OBB validation conf default of 0.01 by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/24561
• Fix stale line anchor for track_high_thresh in track docs by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/24559
• Docs: align markdown table columns by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/24566
• Update CI tests by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/24567
• ultralytics 8.4.53 Retry CUDA backend memory errors during training by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24569

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.52...v8.4.53

🌟 Summary

🚀 Ultralytics 8.4.52 is a major release centered on adding full YOLO26 semantic segmentation support, plus several reliability, export, visualization, and documentation improvements.

📊 Key Changes

• 🆕 New semantic task added across Ultralytics

  • YOLO now supports semantic segmentation as a first-class task alongside detection, instance segmentation, classification, pose, and OBB.
  • This includes support in the CLI, Python API, training, validation, prediction, export, docs, and benchmarks.

• 🧠 New YOLO26 semantic segmentation models

  • Added new pretrained model variants with the -sem suffix, such as:
    • yolo26n-sem.pt
    • yolo26s-sem.pt
    • yolo26m-sem.pt
    • yolo26l-sem.pt
    • yolo26x-sem.pt
  • These models are designed for pixel-level scene understanding, where every pixel is assigned a class.

• 🏗️ Complete semantic segmentation pipeline introduced

  • Added dedicated components for:
    • SemanticSegmentationModel
    • SemanticSegmentationTrainer
    • SemanticSegmentationValidator
    • SemanticSegmentationPredictor
  • This makes semantic segmentation feel like a native YOLO workflow rather than an add-on.

• 🖼️ Semantic masks now appear directly in inference results

  • Added a new SemanticMask result type.
  • Users can now access semantic predictions cleanly in outputs and visualizations.

• 📦 Built-in semantic segmentation datasets added

  • Added dataset configs and docs for:
    • Cityscapes
    • Cityscapes8 for quick testing
    • ADE20K
  • Also added support for PNG mask datasets and even polygon-based semantic training from existing segmentation labels.

• 🔁 Augmentation pipeline upgraded for semantic masks

  • Mosaic, MixUp, CutMix, flipping, perspective transforms, and letterboxing now properly support semantic segmentation masks.
  • This is important for training quality and consistency.

• 📤 Export and deployment support extended

  • Semantic segmentation models now work with export flows and benchmark coverage.
  • Special handling was added for semantic outputs in TensorFlow and multiple export backends.
  • nms=True is automatically disabled for semantic models, since it does not apply to this task.

• 🧪 Much broader testing coverage

  • Added engine, Python, export, benchmark, and dataset tests for the new semantic task.
  • Raspberry Pi skips were added where semantic tests may exceed memory limits.

• 🛠️ Other notable fixes in this release

  • FP16/BF16 FLOPs calculation fixed so half-precision models no longer incorrectly report 0.0 GFLOPs. 📏
  • Final evaluation logging fixed so best metrics no longer overwrite the last epoch in TensorBoard/MLflow. 📈
  • Validation visualizations now respect show_labels and show_conf, making crowded scenes easier to inspect. 👀
  • Dependency handling improved for TensorFlow, Ray Tune, and IMX export by tightening protobuf-related constraints. 📦

🎯 Purpose & Impact

• 🎯 Biggest impact: YOLO26 can now do semantic segmentation natively

  • This opens YOLO to use cases like:
    • autonomous driving
    • road scene understanding
    • medical image parsing
    • land-cover mapping
    • general scene labeling

• ⚡ Easier adoption for existing YOLO users

  • The new task fits into the same familiar YOLO workflow, so users can train and run semantic models with similar commands and APIs they already know.

• 🧩 Broader dataset compatibility

  • Support for both mask-based datasets and polygon labels lowers the barrier to using existing annotations for semantic segmentation projects.

• 📉 More reliable metrics and logging

  • FLOPs reporting is now more accurate for reduced-precision models.
  • Training logs are cleaner and less confusing for experiment tracking tools.

• 🔍 Better debugging and visualization

  • Cleaner validation plots and native semantic result handling make it easier to inspect predictions and analyze model behavior.

• 🚀 Stronger deployment readiness

  • Export, benchmark, and backend updates mean semantic models are not just trainable, but also more practical to validate and deploy.

• 📚 Improved docs and setup guidance

  • New semantic segmentation docs, datasets, CLI references, and install notes make the feature much easier to discover and use.

In short, v8.4.52 is a feature-heavy release led by the launch of YOLO26 semantic segmentation 🎉. For many users, this is the headline change and a meaningful expansion of what Ultralytics can now handle out of the box.

What's Changed

• Fix get_flops dtype mismatch on half-precision models by @ahmet-f-gumustas in https://github.com/ultralytics/ultralytics/pull/24504
• Add https://youtu.be/Jj7WpfiegD0 & https://youtu.be/Kpl6cOUKbb8 to docs by @RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/24514
• fix: 🐞 add googleapis-common-protos dependency to resolve protobuf compatibility issues with TensorFlow and raytune by @onuralpszr in https://github.com/ultralytics/ultralytics/pull/24509
• docs: 📝 add installation instructions for NVIDIA DALI with CUDA 13.x by @onuralpszr in https://github.com/ultralytics/ultralytics/pull/24498
• Update OpenVINO YOLO26 benchmarks with Intel® Core™ Ultra™ X7 358H by @lakshanthad in https://github.com/ultralytics/ultralytics/pull/24512
• Fix custom get_model signature in usage/engine.md by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/24531
• fix: on_fit_epoch_end in final_eval overwriting last epoch metrics by @deependujha in https://github.com/ultralytics/ultralytics/pull/24530
• Fix unrendered links in YOLOv5 compatibility warning by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/24521
• feat: hide labels and conf in visualizations plotting by @K-saif in https://github.com/ultralytics/ultralytics/pull/24305
• Fix docs links with trailing slash with latest zensical 0.0.43 by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/24533
• docs: update enterprise license description by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/24473
• Bump idna from 3.10 to 3.15 in /examples/RTDETR-ONNXRuntime-Python in the pip group across 1 directory by @dependabot[bot] in https://github.com/ultralytics/ultralytics/pull/24540
• ultralytics 8.4.52 YOLO Semantic Segmentation models by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/24518

New Contributors

• @deependujha made their first contribution in https://github.com/ultralytics/ultralytics/pull/24530

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.51...v8.4.52

🌟 Summary

Ultralytics v8.4.51 focuses mainly on better training traceability and clearer deployment/docs updates 📦📝, with the most important change adding the Git commit message to training metadata so models are easier to track, reproduce, and audit.

📊 Key Changes

• Training metadata now includes the Git commit message 🧾
  The headline update from @glenn-jocher adds the current commit subject into:

  • saved checkpoints as git.message
  • Platform training environment metadata as gitCommitMessage
  • Git repository utilities via GitRepo.message

• More robust Git metadata handling 🔧
  Git repository parsing was improved to better read metadata from Git internals, including shared/worktree-style layouts. This helps Ultralytics capture version information more reliably during training.

• Major augmentation pipeline refactor 🛠️
  A substantial internal refactor by @Laughing-q introduced a more unified transform system:

  • BaseTransform now standardizes how image, instance, and semantic-mask transforms are applied
  • augmentations like Mosaic, MixUp, CutMix, CopyPaste, RandomPerspective, RandomFlip, and LetterBox were reorganized around this shared structure

• OpenVINO docs updated with YOLO26 benchmarks 🚀
  The OpenVINO documentation now highlights YOLO26 benchmark results instead of older YOLO11 benchmarks, with refreshed performance data across newer Intel CPUs, GPUs, and NPUs.

• DeepX export documentation expanded 📤
  DeepX was added to the export formats table, with supported export arguments and output folder behavior documented more clearly.

• RT-DETR inference tuning guidance added ⚡
  Docs now explain that users can reduce query count for faster RT-DETR inference, helping users trade a bit of accuracy for lower latency when needed.

• YOLOE export behavior clarified ⚠️
  The docs now clearly warn that exported YOLOE models are static: once exported, prompt-based class configuration is baked into the model and cannot be changed later.

• Ultralytics Platform GPU docs refreshed ☁️
  Platform docs now reflect:

  • more available GPU types
  • new B300 GPU availability
  • updated plan access details
  • revised GPU pricing

• Test/CI compatibility improvement for Axelera export 🧪
  Axelera export tests are now limited to supported PyTorch versions, reducing false failures in CI.

• General documentation cleanup 📚
  Smaller updates include a fixed DeepX link, README simplification, removal of old Weglot docs overrides, and wording/casing polish across docs.

🎯 Purpose & Impact

• Easier experiment tracking and reproducibility 🔍
  Adding the Git commit message makes it much easier to tell what exact code change produced a trained model, especially when many experiments are run close together.

• Better debugging and collaboration 🤝
  Teams using local training or the Ultralytics Platform can now connect checkpoints and cloud runs to a human-readable commit description, not just a hash.

• Stronger foundations for future augmentation work 🧱
  The transform refactor is mostly an internal improvement, but it should make augmentations more consistent, easier to maintain, and safer to extend in future releases.

• Clearer deployment decisions for users 📈
  Updated YOLO26 OpenVINO benchmarks, DeepX export docs, and RT-DETR speed tips help users choose faster deployment settings with more confidence.

• Fewer surprises in export workflows ✅
  The YOLOE warning helps users avoid exporting a model and later discovering that prompt changes no longer work.

• Improved platform transparency 💳
  Updated GPU availability and pricing docs help users better plan cloud training costs and choose the right hardware tier.

Overall, v8.4.51 is less about new end-user model features and more about making training runs easier to understand, reproduce, and deploy reliably 🌟

What's Changed

• Fix broken link by @lakshanthad in https://github.com/ultralytics/ultralytics/pull/24485
• Remove Weglot docs overrides by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24484
• Clean up integrations section in README by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24486
• Scope Axelera export to 2.8.0<=torch<2.12.0 in Tests CI by @lakshanthad in https://github.com/ultralytics/ultralytics/pull/24487
• Add DeepX to export formats table and optimize arg docs by @lakshanthad in https://github.com/ultralytics/ultralytics/pull/24488
• Refactor BaseTransform and unify augmentation pipeline by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/24434
• docs: add RT-DETR query count tip by @artest08 in https://github.com/ultralytics/ultralytics/pull/24466
• Warn that exported YOLOE models are static by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/24496
• docs: normalize casing for PyTorch / NumPy / Matplotlib in prose by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24499
• Update OpenVINO benchmarks with YOLO26 by @lakshanthad in https://github.com/ultralytics/ultralytics/pull/24285
• Update Platform GPU docs for B300 by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24506
• ultralytics 8.4.51 Add Git commit message to training metadata by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24505

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.50...v8.4.51

🌟 Summary

Ultralytics 8.4.50 is mainly a deployment-focused release 🚀, led by a new DeepX export and inference integration that makes it easier to run YOLO models on DeepX NPU edge hardware, along with a few quality-of-life fixes for tuning reliability, mixed-precision model fusion, and RT-DETR documentation.

📊 Key Changes

• New DeepX export support added 🧠⚡
  You can now export Ultralytics models directly with format="deepx" as part of the normal export workflow.

  • Export first converts the model to ONNX, then compiles it into DeepX’s .dxnn format.
  • Export generates a complete deployment folder with:
    • compiled model file
    • config.json for preprocessing and calibration
    • metadata.yaml for model information
  • DeepX export is designed for INT8 quantized deployment and automatically enables int8=True.
  • If no calibration dataset is provided, it defaults to coco128.yaml.
  • Supports standard Ultralytics tasks across modern model families, including detection, segmentation, pose, OBB, and classification.

• New DeepX inference backend added 🔌
  Exported DeepX models can now be loaded back into Ultralytics for inference through a dedicated backend, making deployment more integrated and easier to use.

• DeepX documentation and references added 📘
  This release includes a full DeepX integration guide, reference pages, export docs, and navigation updates so users can more easily learn and adopt the new workflow.

• Hyperparameter tuning now reports failures more honestly 🛠️
  The tuner no longer claims all iterations completed successfully when training runs actually failed.

  • Failed runs are now reflected in the status output
  • Misleading best_hyperparameters.yaml files are no longer written when no valid result exists

• Fix for dtype issues during Conv/BN and Deconv/BN fusion 🎯
  Fusing models in float16 or bfloat16 now works correctly, avoiding dtype mismatch errors during optimization.

• RT-DETR docs improved with eval_idx guidance 📉➡️⚡
  The docs now explain how to reduce RT-DETR decoder layers at inference time to trade a small amount of accuracy for lower latency, without retraining.

🎯 Purpose & Impact

• Biggest impact: easier edge deployment on DeepX hardware 🚀
  This is the headline feature of 8.4.50. Users targeting low-power embedded AI systems can now export YOLO models directly into a format optimized for DeepX NPUs, making deployment much smoother.

• Better out-of-the-box hardware acceleration workflows ⚙️
  By automatically handling INT8 export, calibration setup, and preprocessing config generation, the new DeepX integration reduces manual work and lowers the barrier to specialized edge deployment.

• Improved production readiness for embedded use cases 🤖
  This helps teams building applications like smart cameras, robotics, industrial automation, and other edge AI systems where speed and power efficiency matter.

• More trustworthy tuning results ✅
  Users running hyperparameter search will now get clearer feedback when experiments fail, helping avoid false confidence and bad downstream decisions.

• More stable model optimization in reduced precision 🧩
  The fusion fix is especially useful for advanced users working with float16 or bfloat16, improving reliability in optimized inference pipelines.

• Helpful RT-DETR speed tuning guidance for latency-sensitive users ⏱️
  While this is docs-only, it gives users a practical way to make RT-DETR faster in deployment scenarios without changing training.

Notes

• DeepX export currently requires x86-64 Linux 🐧
• ARM64 export is not supported
• Inference support depends on the DeepX runtime and dx_engine

Overall, 8.4.50 is a strong deployment release, with DeepX integration as the standout addition 🌟—especially valuable for anyone moving Ultralytics models onto specialized edge AI hardware.

What's Changed

• Fix tuner reporting iterations complete when every iteration failed by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/24477
• Docs: add RT-DETE eval_idx instruction by @artest08 in https://github.com/ultralytics/ultralytics/pull/24465
• Fix dtype mismatch in fuse_conv_and_bn and fuse_deconv_and_bn by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/24480
• ultralytics 8.4.50 New Export Integration: Deepx by @ambitious-octopus in https://github.com/ultralytics/ultralytics/pull/23553

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.49...v8.4.50

🌟 Summary

Ultralytics v8.4.49 is mainly a stability and usability release 🛠️—it helps training fail earlier with clearer messages, recovers from corrupted cached weights automatically, and includes several quality-of-life fixes across training, metrics, benchmarking, tuning, tracking, and docs.

📊 Key Changes

• 🚨 Fail-fast training improvements from PR #24478 by @glenn-jocher:

  • Corrupt cached model files now recover automatically: if a cached .pt weight file is damaged or incomplete, Ultralytics now deletes it and re-downloads it instead of failing repeatedly.
  • Clear error when everything is frozen: if freeze is set too high and no trainable parameters remain, training now stops immediately with a direct, understandable error.
  • Clear error for empty classification splits: if class filtering leaves a train/val split with zero images, users now get a helpful FileNotFoundError naming the empty split.

• 🧠 Better distributed training behavior with torch.compile from PR #24413 by @Y-T-G:

  • DDP setup was adjusted to improve compatibility and stability, especially for compiled models and edge cases like batches with no labels.

• 📈 Per-image F1 metric fix from PR #24444 by @fcakyon:

  • Images with no ground-truth objects and no predictions are now scored correctly as a perfect result rather than an F1 of 0.

• 🧪 Hyperparameter tuning randomness fix from PR #24451 by @raimbekovm:

  • The tuner no longer risks repeating the same mutations in very fast iterations, improving search diversity.

• 🛤️ Dataset relative path handling fix from PR #24429 by @auduntorp:

  • Relative paths like ../ are now handled correctly in dataset file lists.

• ⚙️ Benchmarking fix from PR #24450 by @Y-T-G:

  • Prevents the data argument from being incorrectly forwarded into export benchmarking flows.

• 🎯 Tracker API cleanup from PR #24457 by @Laughing-q:

  • Removed the unused frame_rate argument from BYTETracker and BOTSORT, making tracker setup simpler.

• 🚀 CI / runner updates from PR #24436 by @glenn-jocher:

  • Updated runner images and improved TensorRT package caching for smoother CI performance.

• 📚 Docs and reference improvements:

  • Added a proper Python API reference landing page, fixed broken links, improved docs formatting, clarified prediction behavior, and polished several guides.

🎯 Purpose & Impact

• 💸 Saves GPU time and debugging effort: the biggest win in this release is that common training failures are caught before they waste compute or crash later with confusing PyTorch errors.
• 🔁 More self-healing behavior: corrupted cached weights no longer create a “stuck” failure state across repeated runs.
• 🧩 Better error messages for real-world dataset issues: users should spend less time guessing whether a problem is caused by freezing too much, bad caches, or empty class-filtered datasets.
• 🤖 More reliable advanced training setups: users training with DDP, torch.compile, or hyperparameter evolution should see fewer odd edge-case failures.
• 📊 More accurate evaluation reporting: the F1 fix makes validation metrics more fair in empty-image cases.
• 🛠️ Small but practical developer improvements: cleaner tracker initialization, safer benchmarking behavior, and better dataset path handling reduce friction in day-to-day use.
• 📘 Smoother docs experience: while not a model update, the documentation cleanup should make navigation and setup easier for everyone.

Overall, v8.4.49 is less about new models and more about making Ultralytics training and evaluation more robust, more predictable, and easier to troubleshoot ✅

What's Changed

• Support relative paths including ../ by @auduntorp in https://github.com/ultralytics/ultralytics/pull/24429
• Update runner images and cache TensorRT by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24436
• Use static_graph=True for compile and remove find_unused_parameters by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/24413
• docs: Add configurable gradient clipping example to custom trainer guide by @artest08 in https://github.com/ultralytics/ultralytics/pull/24420
• docs: Add Reference section landing page (/reference/index.md) by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24443
• Docs: clarify rect and tuple imgsz for predict (#24440) by @AffanBinFaisal in https://github.com/ultralytics/ultralytics/pull/24441
• Fix broken docs links by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24448
• Fix duplicate data argument in benchmark call by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/24450
• Fix per-image F1 for empty-GT empty-pred images by @fcakyon in https://github.com/ultralytics/ultralytics/pull/24444
• docs: Add RT-DETR variant to per-layer learning rates guide by @artest08 in https://github.com/ultralytics/ultralytics/pull/24419
• docs: fix banner, admonition image, and stacked badges by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24456
• Bump urllib3 from 2.6.3 to 2.7.0 in /examples/RTDETR-ONNXRuntime-Python in the pip group across 1 directory by @dependabot[bot] in https://github.com/ultralytics/ultralytics/pull/24460
• Update URLs to follow 301 redirects to final destinations by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24462
• Professionalize YOLOv5 tutorials by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24468
• Fix Tuner._mutate RNG collisions for sub-second iterations by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/24451
• Remove unused frame_rate in BYTETracker/BOTSORT by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/24457
• Fix docs language banner links by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24474
• docs: add export install tip for benchmarks by @banu4prasad in https://github.com/ultralytics/ultralytics/pull/24464
• Remove flaky DOI badges from docs badge arrays by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24476
• ultralytics 8.4.49 Fail fast on three common training failure modes by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24478

New Contributors

• @AffanBinFaisal made their first contribution in https://github.com/ultralytics/ultralytics/pull/24441
• @auduntorp made their first contribution in https://github.com/ultralytics/ultralytics/pull/24429

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.48...v8.4.49

🌟 Summary

Ultralytics v8.4.48 is a stability-focused release that mainly improves training reliability on Ultralytics Platform and makes failure cases much clearer, with supporting fixes for benchmarking and reporting. 🛠️✅

📊 Key Changes

• (Top priority) Platform training edge-case fixes from PR #24431 by @glenn-jocher:

  • Added guards for empty semantic-mask batches to prevent crashes in segmentation workflows. 🎭
  • Added safe handling for empty RLE keypoint masks in loss computation to avoid invalid operations. 🤸
  • Training now fails with a clear error if no checkpoint (best/last) is saved, instead of failing later with confusing behavior. 💾
  • Model loading now throws a clearer error if a checkpoint depends on a missing ultralytics.* module, with guidance to retrain/use current official models. 📦

• Benchmark/export robustness fix from PR #24418 by @lakshanthad:

  • data is now only passed to export formats that actually support it, reducing export benchmark failures across formats. 📊

• Platform metrics correctness fix from PR #24425 by @mykolaxboiko:

  • training_complete.bestEpoch now reports the real best epoch (from early stopping), not just the final epoch trained. 🎯

• Docs improvement for multi-GPU training from PR #24422 by @artest08:

  • Added a custom trainer guide section showing how to use SyncBatchNorm safely in multi-GPU setups (especially helpful for small per-GPU batch sizes). 📚🧠

• Maintenance/docs housekeeping

  • Dependabot pip updates switched from daily to monthly (PR #24411 by @glenn-jocher). 📅
  • CI docs updated with corrected iOS App Store link and table cleanup (PR #24430 by @glenn-jocher). 📱

🎯 Purpose & Impact

• More reliable training pipelines on edge cases (empty masks, empty keypoint visibility), especially important for automated or mixed datasets. 🚀
• Faster debugging and clearer failures when training output/checkpoint loading goes wrong. This saves developer and MLOps time. ⏱️
• More trustworthy Platform reporting with accurate bestEpoch, improving experiment tracking and decision-making. 📈
• Better cross-format benchmarking stability, so export testing is less brittle. 🔄
• Overall, this release is less about new model features and more about making existing YOLO workflows safer, clearer, and production-friendlier. ✅

What's Changed

• Run Python dependency updates monthly by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24411
• Fix benchmarks passing unsupported data arg to export formats by @lakshanthad in https://github.com/ultralytics/ultralytics/pull/24418
• Fix Platform bestEpoch reporting wrong value at training_complete by @mykolaxboiko in https://github.com/ultralytics/ultralytics/pull/24425
• docs: Add SyncBatchNorm example to custom trainer guide by @artest08 in https://github.com/ultralytics/ultralytics/pull/24422
• Add iOS App Store link to CI docs by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24430
• ultralytics 8.4.48 Fix Platform training edge cases by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24431

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.47...v8.4.48

🌟 Summary

Ultralytics v8.4.47 is a reliability-focused release that fixes a key CLI bug for heatmap colormaps 🎨, while also improving RT-DETR post-processing, remote checkpoint loading, Edge TPU export behavior, and several stability edge cases across loaders and I/O ✅.

📊 Key Changes

• 🔥 Most important (current PR #24219 by @raimbekovm): CLI colormap parsing fixed for Solutions Heatmap

  • Commands like colormap=cv2.COLORMAP_INFERNO now work as documented.
  • Previously, these values were passed as raw strings and could fail in OpenCV.
  • Parser now safely resolves allowed cv2.<UPPERCASE_CONSTANT> values to numeric constants, with security protections still in place 🛡️.

• RT-DETR got a meaningful post-processing upgrade (PR #24403 by @artest08)

  • Top-k selection moved into the decoder head (more aligned with DETR-style behavior).
  • Inference output format is now cleaner and standardized as [cx, cy, w, h, score, class].
  • Predict/val code is simpler and more consistent.

• Edge TPU export is now more robust and explicit (PR #24383 by @lakshanthad)

  • If exporting to Edge TPU without int8=True, Ultralytics now warns and auto-enables it early.
  • Reduces confusing late-stage failures during export.

• Remote checkpoint URI support improved (PR #24395 by @glenn-jocher)

  • Centralized supported remote prefixes and expanded consistency across loading paths.
  • pretrained=ul://... now resolves like other remote model references 🌐.

• Multiple bug fixes across data loading, hub session, and solutions (PR #24397 by @glenn-jocher)

  • Fixed loader counter initialization and tensor path generation bugs.
  • Fixed potential unbound variable in file speed checks.
  • Fixed HUB warning message formatting.
  • Minor analytics/heatmap cleanup for correctness and efficiency.

• Safer image reading behavior (PR #24406 by @Y-T-G)

  • imread() now returns None on missing/unreadable files (matching common cv2.imread expectations), instead of crashing.

• YOLOE validation guardrail added (PR #23982 by @ahmet-f-gumustas)

  • Warns when refer-data class names don’t match eval dataset class names, helping catch silent validation mistakes ⚠️.

• Results serialization improvements (PR #23909 by @glenn-jocher)

  • Mask/keypoint summary values are normalized to Python floats and rounded cleanly for better JSON/API compatibility.

• CI maintenance update (PR #24396 by @UltralyticsAssistant)

  • Slack GitHub Action bumped from v3.0.2 to v3.0.3 in workflows.

🎯 Purpose & Impact

• For CLI users of yolo solutions heatmap: documented colormap syntax now behaves correctly out of the box 🎉.
• For deployment workflows: Edge TPU exports and remote weight loading are more predictable and less error-prone.
• For RT-DETR users: post-processing is more model-native and maintainable, with potential quality/selection behavior benefits.
• For production pipelines: better failure handling (imread), cleaner serialization, and several bug fixes reduce runtime surprises.
• Overall: this release is less about flashy new UI and more about making Ultralytics tooling more trustworthy, consistent, and developer-friendly 🚀.

What's Changed

• Fix Edge TPU INT8 calibration when data arg is not passed by @lakshanthad in https://github.com/ultralytics/ultralytics/pull/24383
• Bump slackapi/slack-github-action from v3.0.2 to v3.0.3 in /.github/workflows by @UltralyticsAssistant in https://github.com/ultralytics/ultralytics/pull/24396
• Fix latent bugs across loaders, hub session, and solutions by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24397
• Resolve platform URIs for checkpoint loading by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24395
• Add refer data and eval dataset class name consistency check for YOLOE validation by @ahmet-f-gumustas in https://github.com/ultralytics/ultralytics/pull/23982
• Return None in imread on read failure to match cv2.imread by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/24406
• Add top-k post-processing to RT-DETR decoder head by @artest08 in https://github.com/ultralytics/ultralytics/pull/24403
• Truncate results to 5 decimals by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/23909
• ultralytics 8.4.47 Fix CLI parsing for solution colormap values by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/24219

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.46...v8.4.47

🌟 Summary

Ultralytics v8.4.46 is a stability-focused release that primarily fixes a key multi-scale training edge case 🔧, while also improving export reliability, hardware support clarity, and documentation quality across YOLO workflows.

📊 Key Changes

• 🚨 Priority fix (PR #24394 by @glenn-jocher): Multi-scale training minimum size clamp

  • Multi-scale random resizing now enforces a safe lower bound of at least one model stride.
  • Added a regression test to prevent future breakage when training with aggressive shrinking on small images.
  • Version bumped from 8.4.45 → 8.4.46.

• 🔁 Resume training safety fix (PR #24386 by @lmycross)

  • Resuming from an already finished checkpoint now correctly stops with “nothing to resume,” instead of running an extra epoch.

• 📦 Export usability and correctness improvements

  • Export completion logs now show the exact artifact path (not just parent folder) (PR #24316).
  • Export argument support was aligned across code/docs for several formats, especially around data and fraction for INT8 calibration (PR #24382).

• 🧠 RKNN support clarified and enforced (PR #24384 by @lakshanthad)

  • RKNN export now clearly blocks INT8-only Rockchip targets (rv1103, rv1106, rv1103b, rv1106b) with a helpful error.
  • Docs now explicitly reflect current RKNN export behavior: FP16-supported targets only.

• 📚 Docs and content polish

  • Fixed strict docs validation warnings and anchor issues (PR #24389).
  • Refreshed tutorial videos/captions in dataset and Streamlit docs for clearer onboarding (PR #24370 + related doc updates).

🎯 Purpose & Impact

• More robust training 🛡️
  The multi-scale clamp fix prevents invalid tiny image sizes, reducing crash risk and instability during training—especially for small-image or aggressive augmentation setups.

• Less wasted compute/time ⏱️
  The resume fix avoids accidental extra epochs when training is already complete.

• Cleaner deployment experience 🚀
  Better export path reporting and clearer per-format argument support make exporting easier to automate and debug.

• Fewer hardware surprises 🤝
  RKNN users now get immediate, explicit feedback on unsupported INT8-only chips, avoiding confusing late-stage failures.

• Better docs trustworthiness ✅
  Improved documentation consistency helps both new and advanced users move faster with fewer mismatches between docs and actual behavior.

What's Changed

• Improve RKNN exports to highlight non-int8 supported chips by @lakshanthad in https://github.com/ultralytics/ultralytics/pull/24384
• Fix: stop resume when training is already finished by @lmycross in https://github.com/ultralytics/ultralytics/pull/24386
• Add https://youtu.be/FvWl00sD4rc to docs by @RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/24370
• Improve data and fraction args across export formats by @lakshanthad in https://github.com/ultralytics/ultralytics/pull/24382
• Improve export results path to point to actual artifact path by @lakshanthad in https://github.com/ultralytics/ultralytics/pull/24316
• Fix docs strict validation warnings by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24389
• ultralytics 8.4.46 Fix multiscale minimum train size by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24394

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.45...v8.4.46

🌟 Summary (single-line synopsis)

Ultralytics v8.4.45 is a hotfix release that primarily restores correct pretrained training behavior for .pt models (like YOLO26 checkpoints), plus important TensorRT/Jetson export reliability improvements and clearer docs. 🚀

📊 Key Changes

• 🔥 Critical training hotfix (PR #24378 by @glenn-jocher):

  • Restored default behavior so YOLO("*.pt").train() and CLI training correctly start from pretrained checkpoint weights.
  • pretrained=False now only disables weight loading when explicitly set.
  • Custom pretrained weight paths are preserved.
  • Classification scratch-training reset behavior was restored.
  • Version bumped to 8.4.45.

• ⚙️ TensorRT compatibility hardening:

  • Blocked problematic TensorRT 10.2.0 (replacing older 10.1.0 exclusion) to avoid known runtime/export issues.
  • Unified TensorRT dependency checks using shared check_tensorrt() for more consistent setup.

• 🤖 Jetson / INT8 export improvements (PR #24368):

  • JetPack 6 Docker image now upgrades TensorRT from 10.3 to 10.7 to fix INT8 + end2end build issues.
  • Automatic end2end disable is now scoped to the specific problematic case: JetPack 6 + TensorRT 10.3.0 + INT8.
  • Added clearer warning messaging and troubleshooting docs.

• 📦 Export argument support expanded:

  • Added data as a valid argument for more export formats (including TensorRT, OpenVINO, CoreML, TFLite, TF.js, MNN, IMX), improving dataset-aware INT8/export workflows.

• 📚 Documentation updates:

  • Hyperparameter tuning guide updated to match actual tuner behavior and output formats.
  • New Ultralytics Platform dataset Clustering docs added (interactive 2D exploration for clusters, outliers, duplicates).

🎯 Purpose & Impact

• ✅ Biggest user impact: fixes a regression where training from .pt checkpoints could silently skip pretrained weights, which could reduce accuracy and waste training time.
• 🧠 More predictable training behavior: defaults now work as users expect, while still allowing explicit scratch training via pretrained=False.
• 🛡️ Fewer deployment pitfalls: better TensorRT version guardrails reduce hard-to-debug export/runtime errors.
• 🚀 Better Jetson experience: INT8 export is more reliable on JetPack 6, with a clear upgrade path for full end2end behavior.
• 🧰 Cleaner export workflows: broader data argument support reduces validation friction across backends.
• 📖 Lower confusion: docs now better reflect real behavior, helping both new and advanced users work faster.

What's Changed

• Exclude tensorrt==10.2.0 to fix libnvinfer_builder_resource_win.so.10.2.0 error by @lakshanthad in https://github.com/ultralytics/ultralytics/pull/24367
• Auto-update tensorrt from 10.3 to 10.7 on JetPack 6 systems to fix YOLO26 int8 build issues by @lakshanthad in https://github.com/ultralytics/ultralytics/pull/24368
• Update hyperparameter tuning guide to match current tuner behaviour by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/24372
• Add data argument to valid args for INT8 export formats by @lakshanthad in https://github.com/ultralytics/ultralytics/pull/24362
• Docs: Datasets Clustering by @sergiuwaxmann in https://github.com/ultralytics/ultralytics/pull/24376
• ultralytics 8.4.44 Apply pretrained arg across model trainers by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24374
• ultralytics 8.4.45 Fix pretrained checkpoint training regression by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24378

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.43...v8.4.45

🌟 Summary

Ultralytics v8.4.44 improves training control and deployment reliability, led by a key fix that makes the pretrained setting behave consistently across trainers 🔧🚀.

📊 Key Changes

• (Top priority) pretrained argument now works consistently in training flows (@glenn-jocher) ✅

  • pretrained=False is now properly respected even when training from a loaded .pt checkpoint.
  • You can now pass a custom pretrained weights path (string) to override checkpoint weights while still reusing the checkpoint’s model config.
  • Removed older classification-only reset logic because shared trainer setup now handles this behavior consistently.

• Export improvements for INT8 workflows 📦

  • Added data as a valid export argument for more backends (including TensorRT, OpenVINO, CoreML, TFLite, TF.js, MNN, IMX, and others), reducing “unsupported argument” friction.
  • Better support for dataset-aware calibration/export pipelines.

• Jetson + TensorRT stability updates 🤖

  • JetPack 6 Docker now upgrades TensorRT from 10.3 to 10.7 to fix YOLO26 INT8 build issues.
  • end2end auto-disable logic for INT8 export is now narrowly targeted to the known problematic combo (JetPack 6 + TensorRT 10.3.0), instead of broader disabling.

• TensorRT version guardrails updated 🛡️

  • Blocks problematic TensorRT 10.2.0 (replacing older 10.1.0 exclusion) in loading/export checks.
  • Centralized TensorRT checks for more consistent setup behavior.

• Documentation updates 📚

  • Hyperparameter tuning guide now matches actual tuner behavior (including crossover usage and clearer fitness/result interpretation).
  • New Ultralytics Platform dataset Clustering docs: interactive 2D similarity view for finding duplicates, outliers, and data patterns faster.

🎯 Purpose & Impact

• More predictable training outcomes 🎯
  Users can trust that pretrained=False truly starts from random initialization, and pretrained="path/to/weights.pt" is honored correctly.

• Easier experiment control and reproducibility 🔁
  Reusing checkpoint configs while controlling weight initialization makes transfer learning and ablation experiments cleaner and less error-prone.

• Smoother deployment on edge devices ⚡
  Jetson/TensorRT fixes reduce export failures, especially for INT8 YOLO26 workflows.

• Fewer environment-related surprises 🧩
  TensorRT version protections help avoid known broken combinations before they cause runtime/export errors.

• Better usability for a broad audience 🌍
  Clearer docs and improved Platform dataset tooling (clustering) help both advanced practitioners and newer users work faster with higher confidence.

What's Changed

• Exclude tensorrt==10.2.0 to fix libnvinfer_builder_resource_win.so.10.2.0 error by @lakshanthad in https://github.com/ultralytics/ultralytics/pull/24367
• Auto-update tensorrt from 10.3 to 10.7 on JetPack 6 systems to fix YOLO26 int8 build issues by @lakshanthad in https://github.com/ultralytics/ultralytics/pull/24368
• Update hyperparameter tuning guide to match current tuner behaviour by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/24372
• Add data argument to valid args for INT8 export formats by @lakshanthad in https://github.com/ultralytics/ultralytics/pull/24362
• Docs: Datasets Clustering by @sergiuwaxmann in https://github.com/ultralytics/ultralytics/pull/24376
• ultralytics 8.4.44 Apply pretrained arg across model trainers by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24374

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.43...v8.4.44

🌟 Summary

Ultralytics v8.4.43 mainly improves dataset reliability by fixing how NDJSON and Ultralytics Platform (ul://.../datasets/...) datasets are resolved during validation, not just training ✅🚀

📊 Key Changes

• Major fix (PR #24365 by @glenn-jocher): unified NDJSON/Platform dataset handling

  • Added a shared resolver: convert_ndjson_to_yolo_if_needed() in ultralytics.data.utils.
  • Replaced trainer’s previous inline, train-only conversion with this shared function.
  • Applied the same conversion path in the validator so standalone validation now also resolves NDJSON and ul:// datasets correctly.
  • Added distributed-safe handling in validation flow to avoid race issues in multi-GPU/distributed runs. 🧩⚙️
  • Patch version bumped from 8.4.42 → 8.4.43.

• Docs update

  • New utility function is now documented in the data utils reference. 📚

• Minor cleanup (PR #24351 by @ahmet-f-gumustas)

  • Replaced redundant for ... in dict.keys() loops with direct dict iteration across several files.
  • No functional behavior change; this is a small code-quality/performance polish. 🧹

🎯 Purpose & Impact

• More consistent dataset behavior across train + val
  Users working with NDJSON or Ultralytics Platform datasets now get the same automatic conversion behavior in both stages, reducing surprises. ✅

• Fewer validation-time failures
  The key practical improvement is preventing validation setup issues when data is provided as .ndjson or ul://... URIs. This makes workflows more robust, especially when running validation independently. 🛡️

• Better stability in distributed environments
  The conversion step is now coordinated for multi-process runs, helping avoid duplicate or conflicting setup work. 🖥️🖥️

• Low upgrade risk
  Aside from the dataset-resolution fix, the rest is mostly internal cleanup with no intended user-facing behavior changes. 👍

What's Changed

• Remove redundant .keys() calls in dict iteration by @ahmet-f-gumustas in https://github.com/ultralytics/ultralytics/pull/24351
• ultralytics 8.4.43 Fix NDJSON dataset validation by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24365

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.42...v8.4.43

🌟 Summary

Ultralytics v8.4.42 focuses on more reliable training under GPU memory pressure 🧠💪, plus several stability, security, export, and documentation improvements that make YOLO workflows smoother for both developers and end users.

📊 Key Changes

• Top priority (current PR #24360 by @glenn-jocher): Better OOM recovery during training 🚀

  • When CUDA runs out of memory in early training, temporary tensors (batch, loss, preds) and trainer loss state are now explicitly cleared before retry.
  • Existing auto-batch retry behavior is preserved (no workflow changes for users).

• Security hardening in downloads and path handling 🔐

  • Safer archive extraction now blocks unsafe zip/tar paths and risky members.
  • Safer URL/URI-derived file handling to prevent unsafe path traversal patterns.
  • Replaced unsafe parsing behavior in CUDA test logic and improved related safeguards.

• Export pipeline refactor + reliability fixes 📦

  • Export dependency setup moved from central exporter into per-format modules (cleaner architecture, easier maintenance).
  • CoreML export now skips unnecessary proxy model loading to avoid export crashes on some Apple Silicon setups.
  • Added a new guide for exporting non-YOLO PyTorch models (timm/torchvision/custom) to multiple targets.

• Backend usability improvement ✨

  • Backends can now be called directly like model(x) instead of only model.forward(x).

• Bug fixes affecting real workflows 🛠️

  • Fixed dataset/image loading when folder names contain square brackets (e.g. ./[data]/).
  • Fixed hyperparameter tuning checkpoint copying for multi-dataset runs.
  • Fixed plot_results() crash path when CSV plotting fails.
  • Added safety check for YOLOE head fusion when one2many branches are absent.

• Platform and docs improvements 📚

  • Platform API docs now clearly document segmentation label payloads using segments (not polygon).
  • Jetson guide adds practical memory optimization tips (headless mode, service trimming, show=False, profiling).
  • Updated Axelera integration to stable SDK 1.6.0.
  • Broader docs cleanup: onboarding/order fixes, activity feed path fixes, updated tutorial videos, annotation links, and removed outdated API endpoints.

🎯 Purpose & Impact

• Fewer failed training runs on limited VRAM 🎯
  The main v8.4.42 change helps training recover more reliably from first-epoch OOM events, reducing frustrating restarts.

• Safer defaults for file and dataset handling 🛡️
  Security-focused path and extraction checks reduce risk when working with external files and archives.

• More dependable exports across formats 🌍
  Export refactors and CoreML fixes improve cross-platform conversion stability, especially for deployment-heavy teams.

• Better developer ergonomics 👩‍💻
  Callable backends, cleaner export internals, and bug fixes in tuning/plotting/path-loading improve everyday productivity.

• Clearer onboarding and learning resources for all users 🙌
  Improved docs (including Ultralytics Platform guidance and Jetson memory tips) make setup and adoption easier for newcomers while still useful for advanced users.

What's Changed

• Add https://youtu.be/m-Lc5MXbydg to docs by @RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/24294
• Document Platform segment label payloads by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24296
• Update Axelera SDK to 1.6.0 with YOLO classifier fix by @tamrobb in https://github.com/ultralytics/ultralytics/pull/24278
• docs: 📝 Add memory optimization tips for NVIDIA Jetson devices by @onuralpszr in https://github.com/ultralytics/ultralytics/pull/24299
• Bump slackapi/slack-github-action from v3.0.1 to v3.0.2 in /.github/workflows by @UltralyticsAssistant in https://github.com/ultralytics/ultralytics/pull/24309
• Add guide for exporting non-YOLO PyTorch models by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/24277
• Add https://youtu.be/hsimB10D6Y0 & https://youtu.be/BpSuXSUzEYY to docs by @RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/24308
• fix(coreml): skip proxy load during export by @amanharshx in https://github.com/ultralytics/ultralytics/pull/24312
• Remove drifting GPU heuristic table from platform quickstart by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/24303
• Added banner to licensing section by @matt-ultralytics in https://github.com/ultralytics/ultralytics/pull/24295
• Correct activity feed access paths on platform by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/24288
• Correct platform onboarding step order and final step by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/24289
• Remove ghost API endpoints from platform docs by @mykolaxboiko in https://github.com/ultralytics/ultralytics/pull/24111
• Fix missing weights copy in hyperparameter tuning iterations by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/24196
• Bump conda-incubator/setup-miniconda from 3 to 4 in /.github/workflows by @dependabot[bot] in https://github.com/ultralytics/ultralytics/pull/24319
• Add __call__ wrapper for BaseBackend.forward method by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/24311
• Set conda workflow permissions by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24317
• add ultralytics.com/annotate link to docs pages by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/24340
• Remove redundant .keys() call in dict membership test by @tempoo04 in https://github.com/ultralytics/ultralytics/pull/24339
• Fix Snyk code security findings by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24342
• Fix undefined ax variable in plot_results by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/24345
• Fix lychee install workflow by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24347
• Fix: FileNotFoundError when loading datasets from paths containing square brackets by @prashansapkota in https://github.com/ultralytics/ultralytics/pull/24353
• YOLOE: check one2many head exists while fusing prompt embeddings by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/24359
• Refactor: move export package requirements from exporter.py into per-format utility modules by @lakshanthad in https://github.com/ultralytics/ultralytics/pull/24245
• ultralytics 8.4.42 Free tensors before OOM batch retry by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24360

New Contributors

• @prashansapkota made their first contribution in https://github.com/ultralytics/ultralytics/pull/24353
• @tempoo04 made their first contribution in https://github.com/ultralytics/ultralytics/pull/24339
• @tamrobb made their first contribution in https://github.com/ultralytics/ultralytics/pull/24278

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.41...v8.4.42

🌟 Summary

Ultralytics v8.4.41 focuses on a key SAM3 video tracking quality fix (fewer ghost IDs) plus a data pipeline reliability improvement for NDJSON datasets, with a large refresh of docs and Ultralytics Platform guidance 📈✨

📊 Key Changes

• 🎯 Major tracking fix (current PR, #24249 by @Y-T-G):

  • SAM3 tracking now enables masklet confirmation filtering by default to reduce false positive “ghost” object IDs.
  • Tracker keep-alive settings were tightened (init_trk_keep_alive and max_trk_keep_alive from 30 → 10), so stale tracks are removed faster.
  • Improved handling of unconfirmed IDs in single-frame inference output.

• 🗂️ Safer NDJSON dataset conversion (#24290 by @glenn-jocher):

  • Converted datasets now use hash-qualified output folders so concurrent jobs don’t overwrite each other’s cache.
  • Hashing ignores temporary signed-URL query strings, preventing unnecessary redownloads.
  • Better fallback behavior when class_names are missing (infers class count more robustly).

• 📚 Documentation and usability upgrades (multiple PRs):

  • New guides for:
    • Fine-tuning YOLO26 on custom data 🧠
    • NVIDIA DALI GPU preprocessing ⚡
    • Modal serverless quickstart ☁️
  • Clearer model-page notes on architecture-only YAMLs vs released pretrained weights.
  • Added docs for the cfg argument (yolo cfg=...) for reusable custom config files.
  • Broad refresh of Ultralytics Platform docs (account, API behavior, datasets, annotation, training, deployment).

• 🧪 CI benchmark stability updates (#24286):

  • Benchmark runners moved from ubuntu-latest to cpu-latest.
  • Slight OBB threshold adjustment to reduce flaky benchmark failures.

• 📦 Version bump: 8.4.40 → 8.4.41.

🎯 Purpose & Impact

• ✅ Better video tracking trustworthiness: Users working with SAM3 in video workflows should see fewer phantom tracks and cleaner ID continuity.
• 🚀 More responsive tracking behavior: Dead/stale objects are dropped sooner, improving real-time tracking quality.
• 🔒 More reliable data ops at scale: NDJSON conversion is safer for teams running parallel training jobs or repeated exports.
• 🧭 Lower onboarding friction: New docs make cloud training, deployment, fine-tuning, and advanced preprocessing easier for both new and experienced users.
• 🌐 Improved Platform clarity: Better alignment between docs and actual Ultralytics Platform behavior helps avoid confusion in production workflows.

What's Changed

• Clarify architecture-only yamls and historical framing on model pages by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/24233
• Document cfg arg in configuration reference by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/24212
• Add Modal Quickstart Guide by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/23414
• Improve docs platform examples by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/24006
• Add fine-tuning guide for YOLO on custom datasets by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/24164
• Refresh platform docs by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24281
• Add NVIDIA DALI GPU preprocessing guide by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/24102
• Migrate benchmarks CI from ubuntu-latest to cpu-latest by @lakshanthad in https://github.com/ultralytics/ultralytics/pull/24286
• ultralytics 8.4.41 Avoid mutable NDJSON dataset cache collisions by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24290
• ultralytics 8.4.41 Fix SAM3 FP ghost IDs in video tracking by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/24249

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.40...v8.4.41

🌟 Summary

Ultralytics v8.4.40 introduces per-image precision/recall/F1 tracking during validation (led by PR #24089 from @Laughing-q), making it much easier to see exactly which images your model handles well or poorly. 📈🖼️

📊 Key Changes

• New per-image validation metrics added to results:
  • precision, recall, f1, tp, fp, fn for each image.
  • Exposed via metrics.box.image_metrics (and also for seg and pose where applicable). ✅
• Detection validation pipeline updated to store image name and compute image-level stats consistently with validation matching logic. 🔍
• Distributed (multi-GPU) validation support now gathers and merges image_metrics correctly across ranks, so results remain complete in larger training setups. 🧠⚙️
• Metrics classes extended with:
  • image_metrics storage
  • update helpers
  • clear/reset helpers to prevent stale metrics between runs.
• Docs updated across validation/task guides (detect, segment, pose, OBB, insights, custom trainer) with examples showing how to access per-image metrics. 📚
• Version bump:
  • 8.4.39 ➜ 8.4.40 🚀

🎯 Purpose & Impact

• Faster debugging of weak samples: You can now pinpoint problematic images directly instead of relying only on dataset-wide averages. 🎯
• Better dataset curation: Find images causing high false positives/false negatives and decide whether to relabel, augment, or rebalance. 🧹
• More actionable model evaluation: Teams get practical, image-level insight for error analysis and iterative improvement. 🔁
• Reliable at scale: Works cleanly in multi-GPU validation, so enterprise and research workflows benefit too. 🏗️
• Broad usability: Useful for both beginners and advanced users working with YOLO models, especially YOLO26 validation workflows. 🤝

What's Changed

• ultralytics 8.4.40 Per-image Precision and Recall by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/24089

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.39...v8.4.40

🌟 Summary

Ultralytics v8.4.39 is a quality-and-usability release focused on clearer run naming (exp-2), better CLI coverage for Solutions, safer rotated-box training behavior, and broad documentation/platform clarity improvements. 🚀

📊 Key Changes

• (Most important) Dashed run path increments are now default via PR #24193 by @glenn-jocher ✅

  • increment_path() now creates names like exp-2 and results-2.txt (instead of exp2, results2.txt).
  • Added regression tests for both directory and file increment behavior.
  • Updated docs/examples to match the new naming format.

• New CLI parity for Solutions via PR #24251 by @raimbekovm 🧰

  • Added direct CLI commands for:
    • yolo solutions region
    • yolo solutions security
    • yolo solutions parking
  • Docs now include CLI examples for these workflows.
  • Notes:
    • Security email alerts still require Python API authentication.
    • Parking requires a prebuilt JSON of parking zones.

• Rotated OBB training safety fix via PR #24260 by @lmycross 🛠️

  • Prevents unintended in-place modification of gt_bboxes in RotatedTaskAlignedAssigner by cloning before edits.

• Small reliability hardening for security alarm flow 🔒

  • Email sending now checks that a mail server is configured before attempting to send.

• CI workflow resilience improvements via PR #24261 by @glenn-jocher ♻️

  • Retry wrappers added around Docker/registry logins.
  • Link-check retries now use fixed backoff to stay within timeout budgets.

• Documentation and platform clarity updates 📚

  • Explorer references reduced in favor of the Ultralytics Platform for no-code dataset analytics and annotation workflows.
  • Sony IMX500 docs now clearly state supported variants (nano-focused note).
  • Neptune/MLflow integration docs corrected to match actual behavior.
  • Platform docs corrected AP region label to Taiwan (not Hong Kong).
  • Pro plan docs clarified to include 4-role RBAC.
  • Many redirected/outdated links were cleaned up.

🎯 Purpose & Impact

• Cleaner experiment tracking 🗂️
  The exp-2 style is easier to read and more consistent across files/folders, reducing confusion when managing many runs.

• Faster adoption of Solutions from CLI ⚡
  Users can now launch region counting, security monitoring, and parking workflows directly from terminal commands with less setup friction.

• More stable rotated-box training pipelines 🎯
  The OBB fix reduces hidden side effects and hard-to-debug training issues, improving trust in advanced detection workflows.

• Better operational reliability 🧪
  CI retry hardening helps reduce flaky failures in automation and publishing workflows.

• Clearer user guidance and platform onboarding 🌐
  Docs now better reflect current best paths (especially Ultralytics Platform) and reduce ambiguity in integrations, licensing, and region details.

What's Changed

• Fix redirected URLs by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24255
• Add CLI parity for region, security, and parking solutions by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/24251
• Add supported model variants note to sony-imx500 integration page by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/24248
• Fix in-place obb gt_bboxes modification in RotatedTaskAlignedAssigner by @lmycross in https://github.com/ultralytics/ultralytics/pull/24260
• Harden retry behavior in docker and links workflows by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24261
• Fix integrations docs drift for neptune and mlflow by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/24259
• Docs: remove explorer references, redirect to ultralytics platform by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/24262
• Correct AP data region label from Hong Kong to Taiwan by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/24264
• Clarify 4-role rbac availability on pro plan by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/24263
• ultralytics 8.4.39 Use dashed incremented run paths by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24193

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.38...v8.4.39

🌟 Summary

Ultralytics v8.4.38 is a stability-focused release that mainly improves model export reliability and consistency across many deployment formats, with additional fixes for training, tracking, and SAM3 behavior. 🚀

📊 Key Changes

• Top priority (PR #24120 by @Laughing-q): Unified standalone export behavior across CoreML, ONNX, OpenVINO, TensorFlow, TorchScript, ExecuTorch, Axelera, RKNN, IMX, NCNN, MNN, and Paddle. 📦
  • CoreML now uses the model’s real input name (instead of assuming "image"), improving compatibility.
  • Export helper functions were standardized with clearer argument names (output_file, output_dir, etc.).
  • Better support for multi-input export scenarios (notably ONNX/OpenVINO).
  • More robust path handling in TensorFlow/IMX export flows.
• Export robustness fixes beyond the main PR
  • Fixed YOLOE export edge case by skipping incompatible fusion when lrpc is present.
  • Improved dependency auto-install fallback to use python -m pip (works better in restricted environments).
  • RKNN now requires newer supported toolkit versions (rknn-toolkit2>=2.3.2).
  • Axelera install commands now allow prerelease packages (--pre) where needed.
• Training reliability improvements 🏋️
  • Fixed resume for non-end2end models (like classification).
  • Adjusted DDP setup order so stride/image-size checks happen safely before wrapping.
  • Checkpoint saves are now skipped whenever EMA weights contain invalid values (NaN/Inf), preventing bad checkpoints.
• Tracking behavior correction 🎯
  • track_buffer now consistently behaves as a true frame count in ByteTrack/BoT-SORT (no hidden FPS scaling).
• SAM3 fixes 🧠
  • Presence-logit clamp now actually applies (in-place).
  • Text-only prompting no longer injects a dummy geometry token, improving confidence behavior.

🎯 Purpose & Impact

• For deployment users: exports should be more predictable and less fragile across devices and runtimes (especially CoreML/OpenVINO/ONNX pipelines). ✅
• For training users: fewer resume/DDP/checkpoint edge-case failures, especially in classification and multi-GPU setups. ⚙️
• For tracking users: track_buffer now matches what you configure, reducing surprising track drop behavior at non-30 FPS. 🎥
• For SAM3 users: improved numerical stability and better text-only grounding quality. 🔬
• Overall: this release is less about new features and more about making existing workflows safer, clearer, and production-friendlier. 🙌

What's Changed

• docs: add trackzone to solutions cli enum by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/24194
• Fix docs macros type and default drift by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/24195
• Fix train resume for non-end2end models by @fcakyon in https://github.com/ultralytics/ultralytics/pull/24173
• Fix torch and torchvision versions on JetPack 5 Docker by @lakshanthad in https://github.com/ultralytics/ultralytics/pull/24199
• fix: update axelera-devkit installation command to allow prerelease versions by @onuralpszr in https://github.com/ultralytics/ultralytics/pull/24190
• Refactor test_engine.py and add OBB/Pose test coverage by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/24197
• Fix docs drift in export-table for executorch, axelera, imx by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/24189
• Fix solutions-args colormap default drift (JET vs DEEPGREEN) by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/24209
• Fix trainer.stride assignment for DDP training by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/24208
• Add missing validated args to TorchScript, CoreML, ExecuTorch integration pages by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/24207
• Fix SAM3 presence logits clamp being a no-op by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/24213
• Clarify YOLO12 pretrained weights are detect-only in model docs by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/24211
• docs: remove outdated Snyk badge from security documentation and update old links by @onuralpszr in https://github.com/ultralytics/ultralytics/pull/24221
• Update annotation viewer zoom shortcut docs by @laodouya in https://github.com/ultralytics/ultralytics/pull/24214
• Fix Sentry bugs 2026-04-14 by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24220
• Fix table format in annotation.md by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/24232
• Update axelera-runtime installation command to allow prerelease versions by @lakshanthad in https://github.com/ultralytics/ultralytics/pull/24230
• fix: OpenVINO 2026 segfault on conda CI by @onuralpszr in https://github.com/ultralytics/ultralytics/pull/24224
• Clarify tt100k uses 221 annotation categories with 45 trainable classes by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/24234
• Docs: add cli tabs for all 12 yolo solutions by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/24217
• Fix: corrected undocumented scaling of track_buffer in trackers by @TimSchoonbeek in https://github.com/ultralytics/ultralytics/pull/24247
• Check for lrpc attribute before attempting fuse during YOLOE model export by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/24239
• Return the model instance when calling fuse method by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/24246
• SAM3: Skip geometry token when using text prompts by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/24244
• Update solution tests to use cached session assets by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/24237
• ultralytics 8.4.38 Unify args naming for standalone export functions by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/24120

New Contributors

• @TimSchoonbeek made their first contribution in https://github.com/ultralytics/ultralytics/pull/24247

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.37...v8.4.38

🌟 Summary

Ultralytics v8.4.37 is a quality + workflow-focused release: the tag PR itself is a version bump, while the main substance is improved hyperparameter tuning (now NDJSON-based for multi-dataset runs), better handling of class imbalance, stronger training reliability, and clearer docs/UI guidance. 🚀

📊 Key Changes

[!WARNING] The mAP calculation has been revised in this release. Reported mAP may be slightly lower than in previous Ultralytics versions, but now more closely matches pycocotools' COCOEval.

• (Priority note) Current PR #24192 by @glenn-jocher: release tag/version update only (8.4.36 → 8.4.37) 📦

  • No direct runtime/model logic change in this PR alone.

• Major tuning upgrade (PR #24179 by @Laughing-q) 🧠

  • Hyperparameter tuning logs moved from CSV to tune_results.ndjson.
  • Better support for multi-dataset tuning with per-dataset fitness tracking.
  • Plot/output naming updated (e.g., tune_fitness.png), and MongoDB sync now aligns with local NDJSON logs.

• Class imbalance support in training (PR #23565 by @ahmet-f-gumustas) ⚖️

  • Added new cls_pw hyperparameter to weight underrepresented classes more during detection training.
  • Default is off (0.0), so existing behavior stays unchanged unless enabled.

• Training stability + robustness fixes 🛡️

  • First-epoch checkpoint is now saved even if EMA has invalid values early on (PR #24170 by @Laughing-q).
  • Fixed local zip dataset regression in HUB/Platform training (safe_download path handling) (PR #24185 by @glenn-jocher).

• Evaluation and CI reliability improvements ✅

  • compute_ap precision edge-case fix for more robust AP calculation (PR #24175 by @Laughing-q).
  • CI/Docker benchmark threshold updates to reduce false failures.

• Cleaner distributed training logs (PR #24177 by @Laughing-q) 🧹

  • Reduced duplicate model info prints in DDP/multi-process training.

• Documentation and platform UX improvements 📚

  • Correct task-specific .load() weights in segment/OBB docs.
  • OpenVINO links updated for YOLO26 optimization notebooks.
  • Platform docs updated for new Settings > API Keys flow.
  • Quickstart diagram made interactive for easier navigation.

🎯 Purpose & Impact

• For ML practitioners tuning across datasets: much better experiment tracking and analysis thanks to NDJSON + per-dataset records. 📈
• For teams with imbalanced data: cls_pw can improve rare-class learning without changing your whole pipeline. 🎯
• For production/research training: fewer broken runs and better checkpoint safety in edge cases. 🔁
• For CI and benchmarking users: fewer flaky failures from minor threshold drift. 🧪
• For new users: clearer docs and safer examples reduce silent misconfigurations and confusion. 🙌

What's Changed

• Add class weights support for handling class imbalance in training by @ahmet-f-gumustas in https://github.com/ultralytics/ultralytics/pull/23565
• Fix AP calculation precision in compute_ap by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/24175
• Fix: Allow checkpoint save for first epoch even when EMA contains NaN/Inf by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/24170
• Fix duplicate model info print for DDP training by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/24177
• Fix benchmark verbose for Dockerfile-nvidia-arm64 by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/24181
• docs: fix wrong weights in segment and obb .load() examples by @amanharshx in https://github.com/ultralytics/ultralytics/pull/24180
• Label Docs non-text code fences by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24176
• docs: update OpenVINO notebook links to YOLO26 optimization by @easyrider11 in https://github.com/ultralytics/ultralytics/pull/24182
• Simplify engine resume tests by @fcakyon in https://github.com/ultralytics/ultralytics/pull/24183
• Fix HUB training regression for local zip datasets by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24185
• replace journey diagram with interactive workflow graph on platform quickstart by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/24186
• ultralytics 8.4.37 NDJSON-based multidataset hyperparameter tuning by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/24179
• improve trainer callbacks docs with precise descriptions by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/24122
• ultralytics 8.4.37 NDJSON-based multidataset hyperparameter tuning by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24192

New Contributors

• @easyrider11 made their first contribution in https://github.com/ultralytics/ultralytics/pull/24182

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.36...v8.4.37

🌟 Summary

Ultralytics v8.4.36 is a stability-focused release that fixes an important training regression for checkpoint-based workflows (especially Ultralytics Platform/HUB usage), plus several documentation clarifications for Platform, Explorer, and Jetson guides. ✅

📊 Key Changes

• 🧠 Critical training fix in Model.train() (PR #24167 by @glenn-jocher)

  • Restores checkpoint-backed model seeding so if a .pt model is already loaded, training reuses that loaded model directly.
  • Prevents the trainer from incorrectly relying on trainer.args.model (which may be a non-file identifier in Platform/HUB flows).
  • Fix is intentionally scoped to the regression introduced in #23640, without broader behavior changes.
  • Adds a dedicated regression test to ensure this checkpoint path stays reliable.

• 🧪 Regression coverage added

  • New test validates that training reuses the in-memory checkpoint model instead of re-parsing model source.
  • Extra test runs confirm detect/segment/classify and core Python model methods remain healthy.

• 📚 Platform and docs alignment updates (supporting PRs)

  • Platform docs now consistently require ultralytics>=8.4.35 for integration.
  • Clarified Platform behavior around deployment defaults (scale-to-zero, single active instance cap), account routes auth scope, dataset/version/export flow, and training/checkpoint expectations.
  • Explorer docs corrected: Explorer is removed in >=8.3.12, with guidance to pin 8.3.11 if needed.
  • Jetson DeepStream links for older JetPack versions now point to NVIDIA archived docs.

🎯 Purpose & Impact

• 🚀 Biggest impact: checkpoint-based training is reliable again in Platform/HUB-style workflows, reducing failed or confusing training starts.
• 🔒 Lower regression risk: targeted code fix + explicit regression test means better confidence in future releases.
• 🧭 Better user clarity: docs now better match real Platform behavior (auth, deployment limits, version requirements), helping users avoid setup mistakes.
• 👥 Broad audience benefit: developers get more predictable training internals, while non-experts get clearer, more actionable documentation and fewer “why did this break?” moments.

What's Changed

• Platform docs updates by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24157
• Use archived DeepStream 6.0.1 and 6.3 links by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/24162
• fix explorer docs version requirement by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/24165
• ultralytics 8.4.36 Fix platform training checkpoint regression by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24167

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.35...v8.4.36

🌟 Summary

Ultralytics v8.4.35 is a stability-focused release that makes training recovery smarter, dataset caching safer, and inference/runtime behavior more reliable—especially when runs hit NaNs or dataset metadata is inconsistent. 🚀🛡️

📊 Key Changes

• NaN training recovery improved (most important, PR #24154 by @glenn-jocher) 🔁

  • Training now recovers from last_good.pt instead of retrying a potentially corrupted last.pt.
  • Checkpoint saves are now skipped if EMA weights contain NaN/Inf, preventing bad checkpoints from propagating.

• Detection dataset cache reliability upgrades (PR #24154) 🗂️

  • Empty detection caches are no longer written.
  • When labels are corrupt, final errors now include clearer reasons (instead of vague failures).
  • Detection dataset checks now accept a dataset directory directly (not only a YAML path).
  • NDJSON-to-YOLO cached conversions are rebuilt when expected split folders are missing.

• Cleaner logs during runs (PR #24154) 🔕

  • Repeated Platform: Model not found warnings are throttled to reduce noise.

• OpenVINO robustness improvements (PR #24156 by @glenn-jocher) ⚙️

  • Better device fallback logic (uses CPU fallback more safely when needed).
  • On Linux ARM64 CPU, inference is forced to FP32 for improved compatibility/stability.

• Training behavior and compatibility fixes

  • Prevents duplicate pretrained weight loading for YAML-based training configs (PR #23640).
  • W&B resume now continues the original run instead of creating a disconnected new one (PR #24110).
  • Added backward compatibility for older YOLO12 Area Attention checkpoints (PR #24152).
  • Added missing isatty() in console capture wrapper to avoid crashes with some libraries like transformers (PR #24143).
  • SAM3 decoder cache-miss dtype fix to prevent coordinate-generation errors (PR #24145).

• Docs and ecosystem updates 📚

  • FastSAM/MobileSAM docs refreshed with YOLO26 benchmark context and clearer reproducibility notes (PR #24140).
  • SAHI YOLO26 tiled inference docs simplified and modernized (PR #24133).
  • Legacy docs now better direct users toward YOLO26 as current recommended model line (PR #24139).

🎯 Purpose & Impact

• Fewer failed long trainings ⏱️
  If your run encounters NaNs, recovery is now more trustworthy, reducing wasted GPU time and repeated crash loops.

• Less dataset-debug frustration 🧪
  Better cache validation + clearer corrupt-label messages make it faster to diagnose data issues.

• More predictable production behavior 🏭
  OpenVINO and logging adjustments improve runtime stability on edge/ARM setups and reduce noisy warnings.

• Smoother experiment tracking and resuming 📈
  W&B resume continuity helps keep metrics in one place, improving experiment history quality.

• Low-risk upgrade with practical benefits ✅
  This release is mostly about reliability and developer experience rather than new model architectures—great for teams running frequent training/inference pipelines.

What's Changed

• Fix cache miss handling for coordinate retrieval by @ausk in https://github.com/ultralytics/ultralytics/pull/24145
• update FastSAM and MobileSAM docs with YOLO26 benchmarks by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/24140
• Add missing type hints to segment2box and scale_boxes in utils/ops.py by @ahmet-f-gumustas in https://github.com/ultralytics/ultralytics/pull/24127
• Fix outdated YAML.dump usage in benchmark.py by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/24146
• update outdated model references to YOLO26 by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/24139
• Update YOLO26 sahi tiled inference code by @RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/24133
• Add missing isatty() to _ConsoleCapture stream wrapper by @fcakyon in https://github.com/ultralytics/ultralytics/pull/24143
• Fix duplicate pretrained weight loading in Model.train() for YAML models by @artest08 in https://github.com/ultralytics/ultralytics/pull/23640
• Resume WandB run when training resumes by @artest08 in https://github.com/ultralytics/ultralytics/pull/24110
• Force OpenVINO ARM inference to FP32 by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24156
• Fix YOLO12 forward pass with old checkpoints by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/24152
• ultralytics 8.4.35 NaN training recovery and dataset cache improvements by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24154

New Contributors

• @ausk made their first contribution in https://github.com/ultralytics/ultralytics/pull/24145

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.34...v8.4.35

🌟 Summary

Ultralytics v8.4.34 is a tuning and stability-focused release 🚀, led by a major new feature: multi-dataset hyperparameter tuning in one run, plus several important reliability fixes and broad YOLO26 documentation updates.

📊 Key Changes

• 🧠 Major feature (PR #24067 by @Laughing-q): Multi-dataset hyperparameter tuning

  • model.tune() now accepts data as either a single dataset or a list.
  • During each tuning iteration, training runs across each dataset, then combines results.
  • Fitness is averaged across datasets, so tuning decisions reflect overall performance, not just one dataset.
  • Tests were updated to validate this workflow (coco8.yaml + coco8-grayscale.yaml).
  • Version bumped to 8.4.34.

• 🛡️ Training resume stability fix (PR #24085 by @Y-T-G)

  • Prevents loss spikes after resume on small datasets by keeping AdamW’s exp_avg_sq state in FP32.
  • Reduces risk of unstable training when loading checkpoints.

• 🔒 Thread-safe ONNX export (PR #24092 by @glenn-jocher)

  • Added export locking so concurrent threads cannot collide in PyTorch’s global ONNX exporter state.
  • Includes a regression test for parallel export safety.

• ⚙️ Robustness fixes in core runtime

  • DDP cleanup now safely handles command-generation failures (PR #24056 by @nameearly).
  • AAttn fixed for non-divisible dim/num_heads cases to avoid shape/group crashes (PR #24114 by @ZoomZoneZero).
  • crop_mask() now clamps negative coordinates before cropping for safer segmentation postprocessing (PR #24115 by @Y-T-G).
  • draw_specific_kpts() now respects user-provided keypoint index order and handles missing confidence values safely (PR #24099 by @onuralpszr).

• 📚 Documentation and ecosystem refresh (many PRs)

  • Large migration to YOLO26 references and fresh benchmarks across Jetson + SAM docs.
  • Jetson setup improved with missing cuDSS dependency instructions for Torch 2.10.0 (PR #24081 by @lakshanthad).
  • DeepStream version mapping/docs links corrected and expanded for newer JetPack versions (PRs #24141, #24142).
  • Ultralytics Platform docs improved (Smart Annotation with SAM + YOLO, split redistribution UX, account settings clarity, banner/link updates).

🎯 Purpose & Impact

• Better real-world tuning quality 🎯
  Multi-dataset tuning helps teams optimize one model for mixed or varied data domains (for example, color + grayscale, or multiple data sources), improving generalization and reducing overfitting to a single dataset.

• More reliable training and export workflows ✅
  Resume training is more stable, distributed cleanup is safer, and ONNX export is more dependable in threaded environments—especially useful in production pipelines.

• Improved deployment and edge guidance 📱
  Updated YOLO26 Jetson benchmarks and setup docs make edge deployment decisions more current and practical.

• Cleaner user experience in docs and platform onboarding ✨
  Better Smart Annotation and dataset split guidance can reduce setup friction and speed up annotation/training workflows on the Ultralytics Platform.

What's Changed

• Update NVIDIA Jetson AGX Orin benchmarks with YOLO26 by @lakshanthad in https://github.com/ultralytics/ultralytics/pull/24068
• update SAM docs with YOLO26 references and fresh benchmarks by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/24071
• update SAM-2 docs with YOLO26 references and fresh benchmarks by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/24072
• update SAM-3 docs with YOLO26 references and correct model sizes by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/24045
• Add missing cuDSS package to JetPack 6 installation guide by @lakshanthad in https://github.com/ultralytics/ultralytics/pull/24081
• Bump onnx from 1.20.0 to 1.21.0rc1 in /examples/RTDETR-ONNXRuntime-Python in the pip group across 1 directory by @dependabot[bot] in https://github.com/ultralytics/ultralytics/pull/24082
• Remove redundant ray tune test by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/24088
• Docs: Redistribute Splits by @sergiuwaxmann in https://github.com/ultralytics/ultralytics/pull/24087
• Document YOLO models in platform smart annotation by @laodouya in https://github.com/ultralytics/ultralytics/pull/24037
• Docs: Update Banner by @sergiuwaxmann in https://github.com/ultralytics/ultralytics/pull/24090
• Docs: Fix Banner Link by @sergiuwaxmann in https://github.com/ultralytics/ultralytics/pull/24091
• Improve Platform Settings page tab docs and add screenshots by @yogendrasinghx in https://github.com/ultralytics/ultralytics/pull/24060
• Fix loss explosion on resume when training on small dataset by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/24085
• Bump onnx from 1.21.0rc1 to 1.21.0 in /examples/RTDETR-ONNXRuntime-Python in the pip group across 1 directory by @dependabot[bot] in https://github.com/ultralytics/ultralytics/pull/24094
• Fix Sentry bugs 2026-04-01 by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24092
• Prevent DDP resource cleanup issues when generate_ddp_command(self) r… by @nameearly in https://github.com/ultralytics/ultralytics/pull/24056
• Add https://youtu.be/Y7cfNkqSdMg to docs by @RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/24098
• fix: 🐞 draw_specific_kpts to respect user-specified indices order by @onuralpszr in https://github.com/ultralytics/ultralytics/pull/24099
• docs: update synk badge with working one by @onuralpszr in https://github.com/ultralytics/ultralytics/pull/24101
• Update NVIDIA Jetson Orin NX 16GB benchmarks with YOLO26 by @lakshanthad in https://github.com/ultralytics/ultralytics/pull/24118
• Clamp coordinates in crop_mask before cropping by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/24115
• Fix AAttn: resolve shape mismatches and crashes when dim is not divisible by num_heads by @ZoomZoneZero in https://github.com/ultralytics/ultralytics/pull/24114
• Update NVIDIA Jetson Orin Nano Super benchmarks with YOLO26 by @lakshanthad in https://github.com/ultralytics/ultralytics/pull/24097
• docs: update DeepStream 7.1 documentation link for JetPack 6.1 by @onuralpszr in https://github.com/ultralytics/ultralytics/pull/24141
• docs: update DeepStream installation instructions for JetPack 7.1 to include DeepStream 9.0 by @onuralpszr in https://github.com/ultralytics/ultralytics/pull/24142
• ultralytics 8.4.34 Multi-dataset support for hyperparameter tuning by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/24067

New Contributors

• @yogendrasinghx made their first contribution in https://github.com/ultralytics/ultralytics/pull/24060
• @ZoomZoneZero made their first contribution in https://github.com/ultralytics/ultralytics/pull/24114
• @nameearly made their first contribution in https://github.com/ultralytics/ultralytics/pull/24056
• @laodouya made their first contribution in https://github.com/ultralytics/ultralytics/pull/24037

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.33...v8.4.34

🌟 Summary

Ultralytics v8.4.33 focuses on a key training reliability fix for end-to-end YOLO workflows, plus improvements to Ray Tune flexibility and CoreML export stability 🚀

📊 Key Changes

• 🔧 Major fix (current PR #24074 by @Laughing-q): Resume-training for end-to-end models now restores loss state correctly

  • Updated resume_training() to properly reinitialize and sync the model’s loss criterion when loading checkpoints.
  • Restores internal loss update counters so one-to-one and one-to-many training behavior continues correctly after resume.
  • Version bumped from 8.4.32 to 8.4.33.

• 🧠 Ray Tune upgrades (PR #23946 by @lmycross): More search algorithms supported in YOLO26 tuning

  • Added search_alg to model.tune(..., use_ray=True) with options like optuna, hyperopt, bohb, ax, nevergrad, zoopt, random, and more.
  • Improved compatibility handling for algorithms with special search-space needs.
  • Renamed internal trial-count handling to align on iterations (clearer API behavior).
  • BOHB now automatically uses the appropriate scheduler (HyperBandForBOHB) instead of generic defaults.
  • Added docs/tests and improved default search space typing (for example close_mosaic now uses integer sampling).

• 🍎 CoreML export fix (PR #24078 by @glenn-jocher): Better detection export with NMS

  • CoreML pipeline now receives explicit model output shape when exporting detection models with nms=True.
  • Reduces shape-related export/runtime issues on Apple devices.

🎯 Purpose & Impact

• More reliable resumed training ✅
  If training is interrupted and resumed from a checkpoint, end-to-end models now continue with the correct loss progression instead of partially reset behavior. This is the most important user-facing fix in this release.

• Stronger hyperparameter tuning workflows 🎯
  YOLO26 users get more control over search strategy in Ray Tune, making large-scale tuning more adaptable to different infrastructure and optimization preferences.

• Smoother mobile/Apple deployment 📱
  CoreML detection exports with integrated NMS are more robust, lowering chances of deployment-time surprises.

• Bottom line: v8.4.33 is a stability + flexibility release—especially valuable for users resuming long trainings and teams doing advanced automated tuning.

What's Changed

• Add support for multiple ray tune search algorithms by @lmycross in https://github.com/ultralytics/ultralytics/pull/23946
• Pass CoreML output_shape arg by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24078
• ultralytics 8.4.33 Progressive loss train resume fix by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/24074

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.32...v8.4.33

🌟 Summary

v8.4.32 is mainly an Axelera AI export expansion release 🚀—it significantly improves how Ultralytics models (including more tasks) are exported and deployed on Axelera hardware, with supporting docs and usability updates across the Ultralytics ecosystem.

📊 Key Changes

• Major (Current PR #23844): Axelera export pipeline refactor and expansion 🧠⚙️

  • Axelera export logic was moved out of the main exporter into a dedicated utility module: ultralytics/utils/export/axelera.py.
  • Export flow now uses a reusable helper, exposed as torch2axelera (replacing older ONNX-centered flow), making the pipeline cleaner and easier to maintain.
  • Axelera export support was broadened beyond detection-only messaging to cover more YOLO task workflows across YOLOv8, YOLO11, and YOLO26 (with one key limitation: YOLO26 segmentation still requires Voyager SDK workaround).
  • Calibration defaults are now task-aware via new TASK2CALIBRATIONDATA, so exports pick better default calibration datasets automatically.
  • Export restrictions/checks were clarified (Linux support, ARM64 Docker exclusion, Torch 2.8+ requirement).

• Axelera runtime/backend improvements 🔌

  • Axelera backend dependency updated to newer runtime packages (axelera-rt / axelera-devkit 1.6.0rc3 path).
  • Backend now loads and optimizes .axm execution pipeline more directly.

• New practical examples for deployment 📁🎯

  • Added full Axelera Python examples for:
    • YOLO26 pose + tracking
    • YOLO11 segmentation
  • These show real end-to-end runtime pipelines without requiring Ultralytics at inference runtime.

• Documentation upgrades (large part of this release) 📚

  • Axelera integration docs were heavily updated: supported tasks table, Python version support (3.10/3.11/3.12), install steps, known limitations, and benchmark tables.
  • New reference docs added for the Axelera export utility.
  • Ultralytics Platform docs clarified:
    • Multi-task annotation persistence (switching task view does not delete other annotations) ✅
    • Dataset upload formats now clearly include .tar.gz / .tgz and NDJSON.
    • Plan-based dataset upload limits clarified: Free 10 GB, Pro 20 GB, Enterprise 50 GB.
  • Docs UI improvements: updated banner, chat widget bump, and table inline-code overlap fix.

🎯 Purpose & Impact

• For developers:

  • Cleaner architecture and reusable export utilities mean easier maintenance, fewer regressions, and faster iteration on Axelera support. 🛠️
  • Better task-aware calibration defaults reduce manual setup and export friction.

• For edge deployment users (Axelera):

  • More practical support for real workflows across YOLOv8/YOLO11/YOLO26 and multiple tasks improves confidence when moving from training to hardware deployment. 🚀
  • New example pipelines make it much easier to go from model export to production-style inference.

• For Ultralytics Platform users:

  • Clearer upload limits/formats and multi-task annotation behavior reduce confusion and accidental workflow mistakes. 📦
  • Better docs UX (chat/script/style/banner updates) improves day-to-day usability.

• Overall:

  • v8.4.32 is a deployment-focused quality release: the biggest value is stronger, more scalable Axelera integration plus clearer documentation that helps both advanced users and newcomers succeed faster. ✅

What's Changed

• Update chat.js v0.2.8 by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24053
• Fix overlapping text with inline code in docs by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/24054
• Add https://youtu.be/VHGfpOrPh-s to docs by @RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/24055
• Add multi-task annotation storage documentation by @t-hakobyan in https://github.com/ultralytics/ultralytics/pull/24044
• 20GB and 50GB Pro and Enterprise docs updates by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24057
• Tar.gz docs by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24058
• Update docs banner by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/24009
• ultralytics 8.4.32 Expand Axelera exportable models and tasks by @lakshanthad in https://github.com/ultralytics/ultralytics/pull/23844

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.31...v8.4.32

🌟 Summary

Ultralytics v8.4.31 is a reliability-focused release that mainly fixes INT8 export calibration for non-square image sizes (the headline change), while also improving training stability, export maintainability, and documentation for deployment and dataset workflows 🚀

📊 Key Changes

• 🔥 Main update (PR #24028 by @Y-T-G): INT8 calibration now works correctly with non-square imgsz

  • Fixes export calibration for commands like imgsz=640,480 with int8.
  • Affects multiple export targets: OpenVINO, TFLite, TensorRT engine, and IMX.
  • Calibration now uses the correct effective size and preserves rectangular shape during preprocessing (LetterBox behavior fixed).

• 🧩 Export system refactor (PR #23914 by @onuralpszr)

  • Export logic was split from one large file into per-format utility modules (torchscript, openvino, coreml, ncnn, mnn, paddle, rknn, axelera, etc.).
  • No major user-facing CLI change, but big internal cleanup for maintainability.

• 🍎 Apple Silicon stability improvement (PR #24038 by @Y-T-G)

  • More aggressive memory clearing on MPS devices to reduce leak-related OOM issues during train/val.

• 📦 Better auto-batch with multi-scale (PR #24051 by @glenn-jocher)

  • Auto batch-size estimation now accounts for larger effective image sizes when multi_scale is enabled.

• 📚 Docs and usability upgrades

  • New COCO-to-YOLO conversion guide (PR #23930 by @raimbekovm).
  • New YOLO26 training recipe guide (PR #23949 by @raimbekovm).
  • Platform docs now clearly explain video inference support on dedicated endpoints (PR #24029 by @t-hakobyan). 🎥

• 🛠️ CI and environment robustness

  • CI runner migration to ubuntu-latest and Codecov v6 updates.
  • Added PyTorch 2.11.0 + Torchvision 0.26.0 slow-test coverage.
  • EdgeTPU install command improved for non-interactive environments (--no-tty).

• ✅ Dataset conversion validation tightening (PR #24031 by @glenn-jocher)

  • Non-classification NDJSON→YOLO conversion now expects a val split (instead of allowing test as substitute).

🎯 Purpose & Impact

• Most important impact: users exporting INT8 models with rectangular inputs should see far fewer calibration mismatches and export failures ✅
• Deployment confidence improves across common edge/runtime formats (OpenVINO, TFLite, TensorRT, IMX) with non-square pipelines 📈
• Training becomes more reliable on Apple Silicon and with multi-scale auto-batch settings 💪
• Developer velocity increases thanks to cleaner export architecture, which should make future exporter fixes/features faster and safer 🧠
• Onboarding gets easier through clearer guides (especially COCO conversion and YOLO26 training best practices), helping both new and advanced users get productive faster 📘✨

What's Changed

• Inference video support documentation by @t-hakobyan in https://github.com/ultralytics/ultralytics/pull/24029
• Improve val availability in NDJSON to YAML by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24031
• Add COCO to YOLO conversion guide by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/23930
• Add YOLO26 training recipe guide by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/23949
• Add --no-tty flag for edgetpu_compiler installation by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/24032
• Bump codecov/codecov-action from 5 to 6 in /.github/workflows by @dependabot[bot] in https://github.com/ultralytics/ultralytics/pull/24034
• Add PyTorch 2.11.0 to Slow Tests CI by @lakshanthad in https://github.com/ultralytics/ultralytics/pull/24035
• refactor: split Exporter export methods into per-format utility modules by @onuralpszr in https://github.com/ultralytics/ultralytics/pull/23914
• Migrate CI to ubuntu-latest by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24049
• Sync Platform embed card theme by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/24048
• Clear MPS memory more aggressively to prevent leak by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/24038
• Include multi_scale in auto-batch computation by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24051
• ultralytics 8.4.31 INT8 calibration with non-square imgsz by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/24028

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.30...v8.4.31

🌟 Summary

Ultralytics v8.4.30 is a focused stability release that fixes and hardens training resume behavior, making interrupted training runs much more reliable 🔄✅.

📊 Key Changes

• Main update (PR #24027 by @glenn-jocher): Refactored resume logic in trainer.py to correctly restore training arguments from last.pt earlier in the resume flow.
• Better checkpoint arg handling: Resume now loads checkpoint config immediately and rebuilds runtime args from it more consistently.
• Safer dataset fallback: If the dataset path stored in a checkpoint is invalid, Ultralytics now falls back to the current data argument instead of failing unexpectedly.
• Override support preserved: Users can still override practical resume-time settings like imgsz, batch, device, workers, cache, freeze, val, and plots (plus related options).
• Clear augmentation warning retained: If custom Albumentations were used before, a warning reminds users to pass augmentations again when resuming, since they can’t be auto-restored exactly.
• Version bump: Package updated from 8.4.29 ➜ 8.4.30.

🎯 Purpose & Impact

• Purpose: Fix failing resume scenarios and make checkpoint restoration more robust after interruptions/crashes. 🛠️
• For users training large models: Fewer resume-related surprises, especially in long-running jobs on cloud/GPU servers. 🚀
• For teams/pipelines: More predictable restart behavior in automated training workflows (CI/CD, scheduled jobs, preemptible instances). 🤖
• No new model architecture/features in this tag: This release is about reliability and continuity, not new YOLO model capabilities. 📌

What's Changed

• ultralytics 8.4.30 Fix training resume by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24027

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.29...v8.4.30

🌟 Summary

Ultralytics v8.4.29 is mainly a training reliability release 🔧—it makes resume=True much safer and clearer, while also adding a helpful new COCO JSON training guide and small CI/docs/test maintenance updates.

📊 Key Changes

• Major (current PR #24021 by @glenn-jocher): safer training resume flow 🔄✅

  • resume=True now only resumes if the checkpoint truly contains resumable training state (like epoch + optimizer state).
  • If the loaded file is just a model weight file (not a full training checkpoint), Ultralytics now:
    • warns clearly ⚠️
    • and falls back to fresh training instead of silently misbehaving.
  • Trainer-side validation is stricter and explicitly rejects invalid resume checkpoints 🚫.
  • Still supports overriding practical run settings during resume (like imgsz, batch, device, workers, cache, patience, validation flags, etc.) ⚙️.
  • Package version bumped from 8.4.28 to 8.4.29 📦.

• New docs feature (PR #24005 + cleanup in #24022): train directly from COCO JSON 📘

  • Added a full guide for training YOLO models on COCO JSON annotations without converting to YOLO .txt labels.
  • Includes custom dataset/trainer pattern, caching behavior, config examples, and FAQ.
  • Follow-up polish improved formatting/readability and fixed YAML indentation clarity.

• CI benchmark alignment with latest model family (PR #23965) 🚀

  • Docker benchmark workflow now uses YOLO26n instead of YOLO11n, aligning CI with current recommended models.

• Stability and maintenance 🧪

  • Prediction test asset URL switched to a more reliable CDN path for zidane.jpg.
  • Example dependency bumps in RTDETR ONNXRuntime example (onnx, requests).

🎯 Purpose & Impact

• For trainers and ML engineers: fewer resume-related surprises, clearer errors, and safer checkpoint handling during interrupted or restarted runs 🛡️.
• For teams running long jobs: reduces risk of accidentally “resuming” from non-resumable files and wasting compute/time ⏱️.
• For users with COCO datasets: easier path to experiment without annotation conversion pipelines, while keeping JSON as source of truth 🧩.
• For platform/CI reliability: benchmark and test updates improve confidence that container/testing flows reflect current recommended YOLO26 usage 📈.

What's Changed

• Update Docker Benchmarks CI to YOLO26 by @lakshanthad in https://github.com/ultralytics/ultralytics/pull/23965
• add guide for training YOLO on COCO JSON without conversion by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/24005
• Use jsdelivr CDN assets URL for zidane.jpg reliability by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24022
• Bump the pip group across 1 directory with 2 updates by @dependabot[bot] in https://github.com/ultralytics/ultralytics/pull/24024
• ultralytics 8.4.29 Training resume fixes by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24021

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.28...v8.4.29

🌟 Summary

Ultralytics v8.4.28 improves training reliability on small datasets by making autobatch smarter, while also strengthening network robustness, backend efficiency, and docs/test stability. 🚀

📊 Key Changes

• (Most important) Autobatch now respects dataset size 📦
  From PR #24020 by @glenn-jocher: automatic batch-size selection is now capped to the number of training images.

  • Added dataset_size through Trainer.auto_batch(), check_train_batch_size(), and autobatch().
  • Autobatch profiling now skips candidate batch sizes larger than the dataset.
  • Final batch recommendation is clamped as a safety check.

• More reliable dataset/platform networking 🌐🔁
  From PR #24010 by @glenn-jocher:

  • Smarter retries for transient download failures (timeouts, connection issues, rate limiting, server errors).
  • Faster failure on non-retryable issues (like local save errors).
  • Platform URL resolution retries increased (3 → 5) with clearer retry behavior and backoff logging.

• TensorFlow backend import optimization ⚡
  From PR #24017 by @Y-T-G:

  • TensorFlow is now imported only for saved_model and pb formats in the TensorFlow backend loader.
  • Avoids unnecessary dependency loading in unrelated inference paths.

• Test reliability improvement for YOLO26 OBB CI 🧪
  From PR #24011 by @glenn-jocher:

  • Updated a test image URL (boats.jpg) to a fixed jsDelivr CDN path for more consistent test behavior.

• Ultralytics Platform docs updated to match current UI 📘
  From PR #24014 by @mykolaxboiko:

  • Removed outdated Activity tab references in model docs.
  • Expanded onboarding docs to cover all tours and restart options.

🎯 Purpose & Impact

• Better small-dataset training defaults ✅
  Users training on tiny datasets are less likely to get impractical batch recommendations, reducing setup friction and odd edge-case behavior.

• More resilient cloud/data workflows 🛡️
  Dataset downloads and platform URL resolution should fail less often on temporary network issues, especially in CI or unstable network environments.

• Cleaner runtime behavior 🧹
  Conditional TensorFlow imports can reduce startup overhead and avoid avoidable environment-related issues.

• Improved developer experience 👩‍💻
  More stable tests and clearer docs mean smoother onboarding, troubleshooting, and release confidence.

What's Changed

• NDJSON and Platform retries and timeouts improvements by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24010
• Use jsdelivr CDN assets URL for boats.jpg reliability by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24011
• Scope Tensorflow import in TensorflowBackend by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/24017
• Update Platform docs: remove model Activity tab, expand onboarding tours by @mykolaxboiko in https://github.com/ultralytics/ultralytics/pull/24014
• ultralytics 8.4.28 Limit autobatch to dataset size by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24020

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.27...v8.4.28

🌟 Summary

Ultralytics v8.4.27 focuses on more reliable Ultralytics Platform training control (especially cancellation handling) ✅, plus several stability fixes for data conversion, postprocessing alignment, Paddle compatibility, and Docker runtime updates 🚀.

📊 Key Changes

• 🔴 Priority update (PR #24008 by @glenn-jocher): Improved Platform training integration

  • Added a new sanitizer to clean callback payloads before sending them to the Ultralytics Platform.
  • Non-JSON-safe float values are now converted safely, preventing callback encoding failures.
  • Centralized cancellation handling in one shared helper, replacing duplicated logic.
  • Applied this shared cancellation flow to both training startup and epoch-end callback paths.
  • Updated callback reference docs for the new helper functions.

• 🧩 COCO conversion robustness (PR #23998 by @raimbekovm)

  • Fixed convert_coco() crashes when annotations are missing or have empty keypoints/segmentation.
  • Keeps annotation lists in sync to avoid IndexError.

• 🎯 Coordinate/mask alignment fix (PR #23995 by @raimbekovm)

  • Updated pad rounding in scale_coords() and scale_masks() to match LetterBox behavior.
  • Reduces 1-pixel misalignment issues for pose/segmentation outputs.

• 🐼 Paddle dependency safety pin (PR #23997 by @Laughing-q)

  • Tightened PaddlePaddle requirements to <3.3.0 for export/inference paths to avoid known breakages.

• 🐳 Docker runtime refresh (PR #23991 by @glenn-jocher)

  • Base image upgraded to PyTorch 2.11.0 (CUDA 12.8, cuDNN 9 unchanged).

• 📚 Documentation discoverability (PR #23980 by @raimbekovm)

  • Added better cross-links to the YOLO26 End-to-End Detection guide from FAQ and argument docs.

🎯 Purpose & Impact

• More dependable Platform workflows 🛡️
  Fewer training tracking failures and more consistent cancellation behavior in cloud-managed runs on the Ultralytics Platform.

• Better training control UX ⏹️
  Cancel signals are handled more predictably, reducing confusion and wasted compute.

• Fewer dataset conversion interruptions 📦
  Users converting imperfect COCO labels are less likely to hit hard crashes.

• Cleaner prediction geometry 🖼️
  Improved coordinate/mask consistency helps downstream quality for segmentation and pose tasks.

• Safer dependency behavior ⚙️
  Paddle users are protected from problematic 3.3.x versions by default.

• Modernized container baseline 🚀
  Docker users get a newer PyTorch runtime with minimal GPU stack risk.

What's Changed

• Pin paddlepaddle<3.3.0 by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/23997
• fix pad rounding in scale_coords and scale_masks to match LetterBox by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/23995
• add end-to-end detection guide cross-links by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/23980
• FROM pytorch/pytorch:2.11.0-cuda12.8-cudnn9-runtime by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/23991
• fix IndexError in convert_coco with missing keypoints or segmentation by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/23998
• ultralytics 8.4.27 Improved Platform training integration by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/24008

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.26...v8.4.27

🌟 Summary

v8.4.26 focuses on reliability and usability improvements: smarter Platform dataset handling (auto-validation split for NDJSON), more robust Platform URL resolution, and an important FP16 SAM inference crash fix—plus CI and docs polish. 🚀

📊 Key Changes

• ✅ Platform NDJSON auto-split (PR #23990, @glenn-jocher)
  If a dataset has a train split but no val/test, Ultralytics now automatically creates a small validation split from training data instead of failing immediately.

  • Uses a deterministic split approach for consistency.
  • Warns users to still create a manual validation split for best quality.

• 🌐 More reliable Ultralytics Platform URI resolution (PR #23990, @glenn-jocher)
  Improved ul://... resolution with:

  • Retry logic for transient connection issues
  • Better timeout strategy (fast connect timeout + longer read timeout for large dataset processing)
  • Clearer handling for auth, permission, not-found, and processing states

• 🧠 FP16 SAM TinyViT inference crash fix (PR #23780, @Edwin-Kevin)
  Fixed a half-precision inference error in SAM TinyViT models (like mobile_sam.pt) caused by dtype mismatch in cached tensors.

  • Model setup order was adjusted so cached eval-time tensors match FP16 expectations.
  • Added/updated CUDA test coverage with half=True.

• ⚙️ CI simplification and stability updates (mainly PR #23990)

  • Removed dedicated HUB CI job and related manual trigger
  • Standardized many CI runners from cpu-latest to ubuntu-latest
  • Temporarily pinned GPU CI to torch<2.11 until CUDA 13 driver support is ready

• 📚 Platform docs improvements

  • Added embedded onboarding videos to key Platform docs pages (Quickstart, Account, Data, Train, Deploy) via PR #23986 by @RizwanMunawar
  • Updated docs chat script to v0.2.7 via PR #23989 by @glenn-jocher

🎯 Purpose & Impact

• Less friction for dataset conversion 📦
  Users importing NDJSON datasets into YOLO workflows are less likely to hit hard failures when a validation split is missing.

• Better reliability in cloud/remote workflows ☁️
  Platform URI retries and timeout tuning reduce flaky failures, especially with large datasets or temporary network instability.

• More stable FP16 segmentation inference 🎯
  Users running SAM TinyViT in half precision on GPU should see fewer runtime dtype crashes and smoother deployment behavior.

• Cleaner maintenance pipeline for Ultralytics 🛠️
  CI changes reduce complexity and improve consistency, helping keep releases stable.

• Faster onboarding for new users 🎥
  The new Platform tutorial videos make it easier for broad audiences to get started quickly with data, training, and deployment.

What's Changed

• Add Platform videos to docs by @RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/23986
• Fix FP16 inference crash from TinyViT cached bias dtype mismatch by @Edwin-Kevin in https://github.com/ultralytics/ultralytics/pull/23780
• Update chat.js v0.2.7 by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/23989
• ultralytics 8.4.26 Platform NDJSON autosplit by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/23990

New Contributors

• @Edwin-Kevin made their first contribution in https://github.com/ultralytics/ultralytics/pull/23780

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.25...v8.4.26

🌟 Summary (single-line synopsis)

Ultralytics v8.4.25 mainly restores TensorFlow.js export for most users ✅, while also improving training speed paths, visualization behavior, box alignment accuracy, CI reliability, and YOLO26/platform documentation clarity 🚀📚.

📊 Key Changes

• (Top priority) TF.js export restored for supported systems 🔧
  From PR #23985 by @glenn-jocher:

  • Added dependency pin ydf<0.13.0 (non-ARM64) to avoid a TensorFlow/protobuf conflict.
  • Removed the previous hard disable on TF.js export.
  • TF.js export is now allowed again on supported environments (still blocked on ARM64 Linux).
  • Platform checks were clarified: TF.js and CoreML now have separate, cleaner rules.
  • Version bumped to 8.4.25.

• Faster training internals for pose/detection/OBB losses ⚡
  From PRs #23937 and #23966:

  • Replaced Python loops with vectorized tensor ops in loss preprocessing and keypoint batching.
  • Keeps model outputs/behavior consistent, but improves efficiency in heavy training paths.

• Bounding box alignment fix in scale_boxes 🎯
  From PR #23967:

  • Corrected padding math to match LetterBox rounding.
  • Fixes subtle systematic box shifts (pixel-level misalignment) on certain image sizes.

• Visualization improvement: confidence can be shown without class labels 👀
  From PR #23951:

  • show_conf=True now works independently of show_labels.
  • Enables “confidence-only” overlays when class text is hidden.

• Docs and platform guidance expanded (especially YOLO26 end-to-end detection) 📘
  From PRs #23947, #23974, #23975, #23931:

  • Added a full guide for YOLO26 end-to-end, NMS-free detection and migration from YOLO11/YOLOv8.
  • Improved API/reference doc rendering (including proper References sections).
  • Broad Ultralytics Platform docs refresh: clearer API key usage, pricing/GPU tables, onboarding/deploy/help updates.

• CI and runner infrastructure updates 🧪
  From PR #23875 and #23974:

  • More workflows moved to cpu-latest for better consistency/speed.
  • Updated runner Docker images to 2.333.0.

• Prediction telemetry enhancement 📡
  From PR #23948:

  • Prediction events now include inference backend metadata for better analytics/debugging.

🎯 Purpose & Impact

• Big win for web deployment 🌐
  Teams exporting to TF.js can now resume browser/JavaScript workflows in most supported environments.

• Fewer export headaches 🧩
  Dependency pinning reduces breakage from upstream package mismatches, making exports more dependable.

• Potentially faster training iterations ⏱️
  Vectorized loss preprocessing should reduce Python overhead and improve throughput, especially for larger batches.

• Better output quality and trust in predictions ✅
  The box-scaling fix helps prevent annoying small coordinate offsets that can affect visual quality and downstream logic.

• Improved usability for debugging and demos 🛠️
  Confidence-only display and richer backend telemetry make model behavior easier to inspect.

• Smoother onboarding for YOLO26 + Platform users 🚀
  Updated docs lower migration friction and make it easier for both new and advanced users to train/deploy via the Ultralytics Platform.

What's Changed

• Added autobacked events name by @ambitious-octopus in https://github.com/ultralytics/ultralytics/pull/23948
• Use new cpu-latest runner for faster CI tests by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/23875
• Vectorize keypoint batching loop for Pose loss calculation by @ahmet-f-gumustas in https://github.com/ultralytics/ultralytics/pull/23937
• Vectorize preprocess for detect/obb loss calculation by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/23966
• Fix References docstring sections rendering incorrectly in API docs by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/23931
• Add end-to-end detection guide for YOLO26 by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/23947
• Update Dockerfile-runner to v2.333.0 by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/23974
• Update Platform docs with RTX PRO 4500 by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/23975
• Fix pad calculation in scale_boxes to match LetterBox rounding by @a17sol in https://github.com/ultralytics/ultralytics/pull/23967
• Allow show_conf to be used independently of show_labels in visualization by @K-saif in https://github.com/ultralytics/ultralytics/pull/23951
• ultralytics 8.4.25 TensorFlow.js fix pin ydf<0.13.0 by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/23985

New Contributors

• @a17sol made their first contribution in https://github.com/ultralytics/ultralytics/pull/23967

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.24...v8.4.25

🌟 Summary

Ultralytics v8.4.24 improves training reliability and clarity on Ultralytics Platform 🎯, while also aligning tuning defaults and docs with YOLO26 best practices 🚀.

📊 Key Changes

• (Most important) Better Platform training error surfacing 🛠️

  • Platform callback error handling now shows the actual server-side error message when available (instead of generic HTTP text).
  • If a Platform training session fails to register, users now get a clearer message:
    “Training will not be tracked on Platform”.
  • After registration failure, Platform callbacks are disabled to avoid noisy follow-up errors.

• Version bump to 8.4.24 🔖

  • Package version updated from 8.4.23 to 8.4.24.

• Ray Tune search space updated to match modern YOLO26 ranges 📈

  • Important tuning ranges were corrected (like lr0, momentum, box, cls, scale).
  • Added missing parameters such as dfl and close_mosaic.
  • Docs/examples were synced so users tune with realistic defaults.

• YOLO26 naming adopted in Streamlit inference selector 🔄

  • Model options switched from yolo11* to yolo26* in the Streamlit solution UI.

• Rockchip RKNN docs refreshed with YOLO26 benchmarks 📘

  • Benchmark tables now reflect YOLO26 models and newer tested package versions.

• TF.js benchmarking disabled (temporary safeguard) ⚠️

  • TF.js export benchmarking is blocked due to a known protobuf dependency conflict, preventing confusing runtime failures.

• Platform docs and UX docs improvements 🧾

  • Clearer plan/GPU tier messaging (including Pro-only H200/B200 access).
  • Added real screenshots to key Platform docs pages.
  • Minor docs cleanup and formatting consistency fixes.

🎯 Purpose & Impact

• Faster troubleshooting for Platform users ✅
  More precise error messages reduce guesswork and support time when training jobs fail.

• Cleaner failure behavior 🧹
  Disabling callbacks after registration failure prevents repeated warning spam and gives a more stable user experience.

• Better model tuning outcomes 🎯
  Updated Ray Tune ranges help users avoid outdated search spaces that could hurt accuracy, especially with YOLO26 workflows.

• Stronger consistency across product and docs 📚
  YOLO26 naming, benchmark updates, and docs alignment make it easier for users to follow current recommended paths.

• Safer defaults for problematic integrations 🛡️
  TF.js benchmark disablement avoids known dependency traps until upstream compatibility improves.

What's Changed

• Fix Platform cloud deployments docs by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/23939
• Missing screenshots placeholder updates by @t-hakobyan in https://github.com/ultralytics/ultralytics/pull/23932
• Add UTM enabled banner redirect by @RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/23952
• Update Rockchip RKNN benchmarks with YOLO26 by @lakshanthad in https://github.com/ultralytics/ultralytics/pull/23953
• Migrate model prefixes from YOLO11 → YOLO26 by @RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/23954
• docs: remove duplicate Dataset Tabs heading by @amanharshx in https://github.com/ultralytics/ultralytics/pull/23940
• sync Ray Tune search space and docs with native tuner ranges by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/23938
• Scope class badges to reference section by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/23962
• ultralytics 8.4.24 Improve Platform train error surfacing by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/23957

New Contributors

• @t-hakobyan made their first contribution in https://github.com/ultralytics/ultralytics/pull/23932
• @amanharshx made their first contribution in https://github.com/ultralytics/ultralytics/pull/23940

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.23...v8.4.24

🌟 Summary

Ultralytics v8.4.23 is mainly a big under-the-hood inference upgrade 🚀: AutoBackend was fully redesigned into modular backend classes, making multi-format model deployment cleaner, easier to maintain, and more reliable across platforms.

📊 Key Changes

• Major architecture refactor (PR #23790 by @Laughing-q) 🧩

  • AutoBackend moved from one large file to a modular backend system (ultralytics/nn/backends/).
  • Added dedicated backend classes for many runtimes: PyTorch, TorchScript, ONNX/IMX, OpenVINO, TensorRT, TensorFlow/TFLite, Paddle, NCNN, MNN, RKNN, Triton, ExecuTorch, Axelera, and CoreML.
  • Introduced a shared BaseBackend interface to standardize model loading, metadata handling, and forward inference behavior.
  • AutoBackend now uses a single format route + backend map instead of many backend boolean flags.
  • Inference call sites (predict/validate/solutions) were updated to align with the new format-driven backend behavior (channels, format, warmup paths, etc.).
  • Added extensive backend reference docs pages under docs/en/reference/nn/backends/.

• Pose dataset conversion improvement (PR #23921 by @glenn-jocher) 🤸

  • NDJSON pose conversion can now infer missing kpt_shape automatically (when possible), instead of failing immediately.
  • Keeps strict validation for inconsistent or ambiguous keypoint data.

• Docs correctness and usability fixes 📚

  • Fixed broken anchor links in several docs pages (PR #23868).
  • Corrected cfg.md defaults for plots and save to True (PR #23917).
  • Removed outdated TensorRT INT8 “batch doubling” documentation (PR #23911).

• Docs/site maintenance and CI updates 🛠️

  • Updated docs JS dependencies (including chat.js and tablesort) (PR #23908).
  • Added Google Tag Manager to docs template (PR #23776).
  • Migrated print() to centralized LOGGER in docs reference builder (PR #23895).
  • Bumped Slack GitHub Action in workflows from v2.1.1 to v3.0.1 (PR #23910).

🎯 Purpose & Impact

• For users deploying models in different formats 🌍
  This release should improve consistency and stability of inference behavior across backends, with fewer backend-specific surprises.

• For developers and contributors 👩‍💻
  The modular backend design is a big win: easier to extend, debug, and maintain than a monolithic AutoBackend.

• For pose dataset workflows 📦
  NDJSON conversion is more forgiving and practical, reducing friction when metadata is incomplete.

• For documentation trustworthiness ✅
  Multiple fixes make docs more accurate and navigable, reducing confusion around defaults and TensorRT calibration behavior.

Overall, v8.4.23 is a foundational engineering release: not a flashy model change, but a high-impact internal upgrade that strengthens the reliability of the Ultralytics inference stack over time.

What's Changed

• fix broken anchor links in docs by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/23868
• Update mkdocs.yml chat.js @v0.2.5 by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/23908
• replace print() with LOGGER in build_reference.py by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/23895
• Bump slackapi/slack-github-action from 2.1.1 to 3.0.1 in /.github/workflows by @dependabot[bot] in https://github.com/ultralytics/ultralytics/pull/23910
• fix incorrect plots and save defaults in cfg.md by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/23917
• fix outdated INT8 batch doubling references in TensorRT docs by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/23911
• Infer NDJSON Pose keypoint length by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/23921
• Fix shape and type comment bugs in Autobackend and Tuner by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/23918
• added google tag manager to main.html by @matt-ultralytics in https://github.com/ultralytics/ultralytics/pull/23776
• ultralytics 8.4.23 Refactor AutoBackend into modular per-backend classes by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/23790

New Contributors

• @matt-ultralytics made their first contribution in https://github.com/ultralytics/ultralytics/pull/23776

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.22...v8.4.23

🌟 Summary

Ultralytics v8.4.22 focuses on better hardware support and reliability, led by a key new feature: basic Huawei Ascend NPU device support in select_device 🚀.

📊 Key Changes

• 🔥 Major (Current PR #23902 by @GiantAxeWhy): Huawei Ascend NPU parsing added

  • You can now use device=npu or device=npu:0 in Ultralytics.
  • Added checks for torch_npu installation and NPU availability.
  • Validates NPU index and gives clearer errors for invalid inputs.
  • Explicitly blocks unsupported multi-NPU strings like npu:0,1 for now.

• 🧠 Multi-GPU training fix (DDP + Albumentations)

  • Custom Albumentations transforms are now serialized/deserialized safely for DDP subprocesses.
  • Fixes failures where custom augmentations could break in multi-GPU runs.

• 📦 Export stability improvements

  • ONNX opset selection is safer on newer Torch versions (Torch 2.9+ capped more conservatively).
  • TensorRT INT8 exports now auto-disable end2end for incompatible TensorRT versions (<=10.3.0) to prevent known build failures.
  • INT8 calibration now handles small datasets more gracefully by reducing batch size instead of hard-failing in some cases.

• 🐳 Docker and CI environment updates

  • Base Docker image moved to PyTorch 2.10.0 (CUDA/cuDNN runtime unchanged).
  • Added UV_BREAK_SYSTEM_PACKAGES=1 across Dockerfiles for more consistent installs.
  • New Dockerfile-runner-cpu and broader CI runner improvements (CPU labels, tooling like sudo, nodejs, npm, gpg).
  • CI expanded to test Torch 2.10.0 + Torchvision 0.25.0.

• 🛠️ Reliability/bug fixes

  • FastSAM text-prompt index remapping fix (better mask selection consistency).
  • ClearML callback fix to prevent remote override conflicts with Ray Tune hyperparameters.
  • Dataset cache loading/saving made more robust against environment mismatches and partial-write failures.

• 📚 Docs and dataset additions

  • Added DOTA128 OBB dataset config/docs (good middle ground between DOTA8 and full DOTA).
  • Expanded Ultralytics Platform docs (SAM model selection incl. SAM 3, team/workspace docs, training/billing clarifications, Docker image guides).
  • Device docs updated to include NPU usage.

🎯 Purpose & Impact

• For Ascend users: You now get first-class basic device selection for Huawei NPUs, making it easier to run training/inference without custom hacks ⚡.
• For training at scale: Multi-GPU workflows with custom augmentations are more dependable, reducing frustrating DDP issues ✅.
• For deployment/export users: Fewer export-time surprises, especially with ONNX/TensorRT compatibility on modern stacks 📤.
• For Docker/CI users: More up-to-date and consistent environments mean fewer setup/build breakages 🧩.
• For everyone: Overall this release is about practical robustness—better hardware compatibility, safer exports, and smoother day-to-day usage with YOLO11/YOLO26 ecosystems 🙌.

What's Changed

• FROM pytorch/pytorch:2.10.0-cuda12.8-cudnn9-runtime by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/23807
• Add UV_BREAK_SYSTEM_PACKAGES environment variable by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/23809
• Add Torch 2.10 with updated Torch and torchvision versions by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/23808
• Bump docker/setup-buildx-action from 3 to 4 in /.github/workflows by @dependabot[bot] in https://github.com/ultralytics/ultralytics/pull/23813
• Fix index remapping in FastSAM by @orcinus in https://github.com/ultralytics/ultralytics/pull/23764
• Update RUNNER_VERSION to 2.332.0 by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/23826
• Add Dockerfile-runner-cpu by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/23827
• Update Dockerfile-python description and derivatives by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/23828
• Remove Runner SHA checks by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/23829
• Update Dockerfile-runner-cpu with sudo by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/23830
• Add nodejs and npm to runner images by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/23832
• Platform SAM3 docs by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/23834
• Fix callable type hint to Callable in BasePredictor.add_callback() by @ahmet-f-gumustas in https://github.com/ultralytics/ultralytics/pull/23836
• Remove unused task attribute from metrics classes by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/23835
• fix: prevent ClearML task.connect() from overriding Ray Tune hyperparam by @knotgrass in https://github.com/ultralytics/ultralytics/pull/23833
• Fix incorrect link in ONNX integration docs FAQ by @sjsr-0401 in https://github.com/ultralytics/ultralytics/pull/23838
• remove resolved TODO comment from OBB head by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/23848
• Exclude OpenVINO 2026.0.0 from Conda CI by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/23839
• fix: 🐞 Enable Albumentations augmentations for DDP training by serializing and deserializing transforms by @onuralpszr in https://github.com/ultralytics/ultralytics/pull/23849
• Update setuptools requirement from <=81.0.0 to <=82.0.1 by @dependabot[bot] in https://github.com/ultralytics/ultralytics/pull/23846
• Disable end2end for TensorRT INT8 exports with tensorrt<=10.3.0 by @lakshanthad in https://github.com/ultralytics/ultralytics/pull/23872
• docs: 📝 notes for multi-GPU training for custom code by @onuralpszr in https://github.com/ultralytics/ultralytics/pull/23876
• Add gpg package for Dockerfile-runner-cpu by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/23878
• docs: 📝 update Docker image list with new images by @onuralpszr in https://github.com/ultralytics/ultralytics/pull/23884
• Add dota128 dataset by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/23885
• Add ModuleNotFoundError in classify caching check to prevent numpy version conflict by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/23892
• Docs: Fix CoreML export table to include agnostic_nms when nms=True by @K-saif in https://github.com/ultralytics/ultralytics/pull/23867
• Add https://youtu.be/4VTuqfrOIws to docs by @RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/23873
• Remove unused tip from export-table.md by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/23898
• fix build_docs.py portability for local development by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/23883
• Docs reference strict mode by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/23900
• Calibration dataset size < batch fix by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/23899
• fix docs CI: remove unused black dep and fix publish commit message by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/23896
• Update format.yml to cpu-latest by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/23901
• handle prettier failures gracefully in build_reference.py by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/23882
• docs: ⚡ Update Triton Inference Server guide by @onuralpszr in https://github.com/ultralytics/ultralytics/pull/23880
• Update docs.yml to cpu-latest runner by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/23904
• Greyscale test fix by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/23903
• docs(export): add FAQ explaining FP32 output with end2end=True and ha… by @Shreyas-S-809 in https://github.com/ultralytics/ultralytics/pull/23759
• Update local docs build instructions in README by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/23877
• Docs: Platform prelaunch documentation updates by @mykolaxboiko in https://github.com/ultralytics/ultralytics/pull/23888
• Fix screenshot placeholders in Teams and Cloud Training docs by @mykolaxboiko in https://github.com/ultralytics/ultralytics/pull/23905
• ultralytics 8.4.22 Huawei Ascend NPU device parsing in select_device by @GiantAxeWhy in https://github.com/ultralytics/ultralytics/pull/23902

New Contributors

• @sjsr-0401 made their first contribution in https://github.com/ultralytics/ultralytics/pull/23838
• @Shreyas-S-809 made their first contribution in https://github.com/ultralytics/ultralytics/pull/23759
• @mykolaxboiko made their first contribution in https://github.com/ultralytics/ultralytics/pull/23888
• @GiantAxeWhy made their first contribution in https://github.com/ultralytics/ultralytics/pull/23902
• @knotgrass made their first contribution in https://github.com/ultralytics/ultralytics/pull/23833
• @orcinus made their first contribution in https://github.com/ultralytics/ultralytics/pull/23764

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.21...v8.4.22

🌟 Summary

Ultralytics v8.4.21 improves reliability for Rockchip RKNN exports (main change) and also adds better tuning isolation, expanded C++ pose example support, and clearer YOLO26 optimizer guidance. 🚀

📊 Key Changes

• ✅ Main priority (PR #23806 by @Laughing-q): RKNN export path fix

  • RKNN filename generation was refactored to use safer path handling (Path(...).stem) instead of fragile string replacement.
  • Export output naming is now cleaner and more consistent, especially for unusual file names/paths.
  • Version bumped from 8.4.20 → 8.4.21.

• 🧪 Ray Tune reliability improvement (PR #23793 by @Y-T-G)

  • Each hyperparameter trial now resets trainer state (trainer = None) before running.
  • Prevents stale state leakage between trials.

• 🧍 ONNXRuntime C++ example gains pose support for YOLOv8-family pose models (PR #23786 by @chendao12138)

  • Adds pose tensor parsing, keypoint decoding/scaling, NMS for pose results, and visualization.
  • Supports FP32 and FP16 in the example pipeline.
  • Note: YOLO26 is not yet supported in this specific C++ example flow.

• 📘 YOLO26 training docs improved around MuSGD (PRs #23800, #23804 by @deriiinjv and @monkeyjack123)

  • Clearer explanation of when MuSGD is used with optimizer=auto.
  • Better practical guidance on when to try MuSGD vs standard options.

🎯 Purpose & Impact

• For deployment users on Rockchip 📦
  RKNN exports should now fail less often due to naming/path edge cases, making deployment workflows more dependable.

• For ML engineers tuning models 🔁
  Ray Tune trials are more isolated and reproducible, reducing hard-to-debug inconsistencies.

• For C++/edge developers ⚙️
  Easier starting point for pose inference in ONNXRuntime C++ projects using YOLOv8-style pose models.

• For YOLO26 trainers 🧠
  Optimizer behavior is easier to understand, helping users make better training choices faster.

What's Changed

• Reset trainer before each trial with Ray Tune by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/23793
• Add YOLOv8 pose estimation support to ONNXRuntime C++ example by @chendao12138 in https://github.com/ultralytics/ultralytics/pull/23786
• Docs: Add explanation for when MuSGD optimizer is used by @deriiinjv in https://github.com/ultralytics/ultralytics/pull/23800
• docs: clarify MuSGD usage in training optimizer guide by @monkeyjack123 in https://github.com/ultralytics/ultralytics/pull/23804
• ultralytics 8.4.21 Fix Rockchip RKNN export path by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/23806

New Contributors

• @deriiinjv made their first contribution in https://github.com/ultralytics/ultralytics/pull/23800
• @chendao12138 made their first contribution in https://github.com/ultralytics/ultralytics/pull/23786
• @monkeyjack123 made their first contribution in https://github.com/ultralytics/ultralytics/pull/23804

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.20...v8.4.21

🌟 Summary

Ultralytics v8.4.20 is a stability-and-usability release focused on cleaner hyperparameter tuning, more reliable deployment/export workflows, and improved docs for both YOLO models and the Ultralytics Platform 🚀

📊 Key Changes

• (Most important) Ray Tune cleanup in current PR #23772 🧹

  • Removed hardcoded tuner_callbacks and dropped built-in W&B callback wiring from RunConfig in tuning.
  • Version bumped from 8.4.19 → 8.4.20.

• RKNN export reliability improved (#23802) 📦

  • Added ONNX version guard (onnx<1.19.0) and enforced ONNX opset cap (<=19) for RKNN conversion.
  • Explicitly supports smoother export paths for YOLO26 models on Rockchip toolchains.

• FastSAM prompt accuracy fix (#23766) 🎯

  • CLIP prompting now masks non-target neighboring segments before scoring.
  • Reduces false positives in overlapping/contained regions (for example, object-inside-background cases).

• ByteTracker consistency update (#23771) 🛠️

  • Optional score fusion now also applies in second-stage association (when fuse_score=True), aligning behavior across tracking stages.

• Better YAML error messages (#23767) ✅

  • Failed YAML loads now return clearer syntax errors with file context and validation guidance, instead of ambiguous fallback failures.

• Jetson/JetPack 6 stack refresh (#23788 + #23801) 🤖

  • Updated Jetson Docker/runtime stack to newer Torch/Torchvision/ONNX Runtime GPU versions.
  • Compatibility checks now recognize torch 2.10 + torchvision 0.25.

• Docs and platform improvements 📚

  • New SAM vs YOLO segmentation benchmark section includes YOLO26n-seg comparisons (#23782).
  • Platform docs expanded for pose skeleton templates and dataset versioning workflows (#23787, #23796).
  • YOLOE tutorial embed updated (#23798).
  • CI/workflow dependency/auth updates (#23763, #23799).

🎯 Purpose & Impact

• More robust tuning workflows ⚙️

  • By removing tightly coupled callback logic, Ray Tune runs are less fragile in mixed environments (especially distributed tuning setups).

• Fewer deployment surprises 🚀

  • RKNN export safeguards reduce version mismatch errors and manual troubleshooting for edge deployment users.

• Higher output quality in segmentation/tracking 👀

  • FastSAM text prompts become more trustworthy in crowded scenes.
  • ByteTracker behavior is more consistent when confidence fusion is enabled.

• Smoother setup for modern environments 🔧

  • Jetson and PyTorch compatibility updates reduce install/runtime friction for developers upgrading dependencies.

• Better onboarding and reproducibility 🧠

  • Clearer YAML errors help users fix configs faster.
  • Platform dataset versioning and skeleton-template docs improve team workflows and repeatable training.

What's Changed

• Use github.Auth.Token in merge action by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/23763
• Better YAML error introspection by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/23767
• Fix for #23765 FastSAM CLIP prompting by @orcinus in https://github.com/ultralytics/ultralytics/pull/23766
• Apply to ByteTracker second association by @ahmet-f-gumustas in https://github.com/ultralytics/ultralytics/pull/23771
• Update sam-3.md speed comparison by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/23782
• Platform Skeleton templates docs by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/23787
• Update JetPack 6 Docker image with torch 2.10 by @lakshanthad in https://github.com/ultralytics/ultralytics/pull/23788
• Platform dataset versioning docs by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/23796
• Bump docker/login-action from 3 to 4 in /.github/workflows by @dependabot[bot] in https://github.com/ultralytics/ultralytics/pull/23799
• Add https://youtu.be/JcZsqUc8PMM to docs by @RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/23798
• Update torch and torchvision compatibility table to latest by @lakshanthad in https://github.com/ultralytics/ultralytics/pull/23801
• Fix RKNN exports to support YOLO26 models by @lakshanthad in https://github.com/ultralytics/ultralytics/pull/23802
• ultralytics 8.4.20 Remove redundant hardcoded tuner_callbacks by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/23772

New Contributors

• @orcinus made their first contribution in https://github.com/ultralytics/ultralytics/pull/23766

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.19...v8.4.20

🌟 Summary (single-line synopsis)

Ultralytics v8.4.19 focuses on much more reliable Ultralytics Platform training sync (especially model tracking via model_id), plus a few quality improvements for SAM outputs, lightweight model stability, and clearer YOLO26 end-to-end docs 🚀

📊 Key Changes

• 🔴 Most important (current PR #23761 by @glenn-jocher): Platform training model_id fix

  • Training callbacks now consistently pass model_id during model uploads and training events.
  • Platform callback state was refactored into a single trainer.platform context (instead of scattered trainer fields), including:
    • model_id
    • last upload time
    • cancellation status
    • console/system loggers
  • Upload and event flow was tightened for better session continuity, cancellation handling, and final result linking.
  • Platform URL handling now better supports server-adjusted model slugs (for example when names auto-increment).

• 🧠 SAM prediction cleanup (#23751 by @Laughing-q)

  • Added NMS in SAM3SemanticPredictor postprocessing and feature-inference paths.
  • Reduces duplicate overlapping detections and keeps boxes/masks aligned.

• 🛡️ PSA attention edge-case fix (#23758 by @Y-T-G)

  • Prevents zero attention heads in PSABlock for very small channel configs.
  • Improves robustness for lightweight/custom model scales.

• 📘 Docs clarification for end-to-end models (#23720 by @raimbekovm)

  • Clarifies that in end-to-end mode (YOLO26, YOLOv10), max_det and agnostic_nms are supported directly.
  • Notes that setting end2end=False is mainly for enabling traditional iou-based NMS behavior.

• ⚙️ CI maintenance

  • GitHub Actions artifact upload/download actions were bumped to newer major versions (#23748, #23749).
  • African Wildlife dataset tutorial video embed updated in docs (#23757). 🎥

🎯 Purpose & Impact

• For Ultralytics Platform users: ✅
  Training runs should now be more reliable to track end-to-end, with fewer mismatches between uploaded checkpoints and the correct model session. This is the biggest practical improvement in this release.

• For production and team workflows: 📡
  Better callback state management means cleaner event handling, better cancellation behavior, and improved confidence that results and artifacts land in the right place.

• For segmentation users (SAM): 🎯
  Cleaner predictions with fewer duplicate overlaps can improve downstream usability and visual quality.

• For custom/small models: 🧱
  Fewer architecture edge-case failures when using very small widths or low channel counts.

• For all users reading docs: 📚
  Clearer expectations around end-to-end inference/validation args in YOLO26 and YOLOv10 reduce confusion and setup mistakes.

What's Changed

• Document max_det and agnostic_nms support in end2end mode by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/23720
• Bump actions/download-artifact from 7 to 8 in /.github/workflows by @dependabot[bot] in https://github.com/ultralytics/ultralytics/pull/23749
• Bump actions/upload-artifact from 6 to 7 in /.github/workflows by @dependabot[bot] in https://github.com/ultralytics/ultralytics/pull/23748
• Add NMS to SAM3SemanticPredictor to eliminate overlapping boxes by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/23751
• Add https://youtu.be/EXYB-dbgJjY to docs by @RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/23757
• Fix division by zero in PSABlock num_heads by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/23758
• ultralytics 8.4.19 Platform training pass model ID fix by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/23761

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.18...v8.4.19

🌟 Summary

Ultralytics v8.4.18 is a reliability-focused release that improves downloads and dependency installs, while also adding stronger ExecuTorch export support (including Pose) for smoother edge/mobile deployment 🚀.

📊 Key Changes

• 🔧 Priority fix (current PR): safe_download() now handles URLs with spaces by encoding spaces as %20 (PR #23736 by @glenn-jocher).
  • Prevents broken downloads from links like file paths or hosted assets containing spaces.
• 📱 Major export upgrade: ExecuTorch refactor + first-class utility module (PR #23734 by @lakshanthad).
  • New ultralytics/utils/export/executorch.py with reusable torch2executorch() and executorch_wrapper().
  • Export pipeline now uses this shared utility.
  • Added Pose-safe decoding patch for better XNNPACK compatibility.
  • Benchmarks no longer block Pose + ExecuTorch inference.
  • New reference docs page for ExecuTorch export.
• 📦 Dependency handling improvement for git+ requirements with version constraints (PR #23737 by @fcakyon).
  • Fixes invalid install behavior by cleaning git URL requirement strings before install.
• 📒 Notebook/API modernization (PR #23745 by @ahmet-f-gumustas).
  • Object tracking notebook updated from deprecated Annotator.seg_bbox() to modern Results.plot(color_mode="instance").
• 📚 Docs updates
  • OBB docs now point to a newer YOLO26 tutorial video (PR #23738 by @RizwanMunawar).
  • Vercel Web Analytics added to docs site template for usage insights (PR #23746 by @glenn-jocher).

🎯 Purpose & Impact

• More reliable downloads for everyone ✅
  The safe_download() fix is small but high-impact: fewer random failures when fetching models/assets from imperfect URLs.
• Better edge deployment path 📲
  ExecuTorch export is now cleaner, more maintainable, and better for Pose workflows—helpful for teams deploying YOLO26 on mobile/embedded targets.
• Fewer setup headaches 🧩
  Git-based dependency auto-update now works more as users expect, reducing installation friction in custom environments.
• Examples stay working and easier to follow 🎥
  Notebook/API updates reduce breakage from old methods and keep learning materials aligned with current Ultralytics usage.
• Docs quality can improve faster over time 📈
  Analytics and refreshed tutorials help the team prioritize what users need most.

What's Changed

• Add version constraint support for git URL requirements by @fcakyon in https://github.com/ultralytics/ultralytics/pull/23737
• Refactor ExecuTorch export to executorch.py by @lakshanthad in https://github.com/ultralytics/ultralytics/pull/23734
• Fix outdated Annotator.seg_bbox in object tracking notebook by @ahmet-f-gumustas in https://github.com/ultralytics/ultralytics/pull/23745
• Add https://youtu.be/128JhhR2DlM to docs by @RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/23738
• Docs analytics by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/23746
• ultralytics 8.4.18 safe_download() URLs with spaces by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/23736

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.17...v8.4.18

🌟 Summary (single-line synopsis)

Ultralytics v8.4.17 makes NDJSON dataset conversions resplit-friendly—reusing existing images, cleaning stale labels, and avoiding unnecessary downloads for faster iteration 🚀📦

📊 Key Changes

• NDJSON dataset re-split support (priority change) ♻️🧹
  • Detects when the dataset output folder already exists and reuses previously downloaded images when you change train/val/test splits.
  • For non-classification tasks (detect/segment/pose), removes the existing labels/ directory before reconversion to prevent stale annotations.
  • If an image is missing from the current split, it will move (rename) it from another split (train/val/test) before downloading from the URL.
  • Adds a content hash (ignoring signed/rotating URLs) stored in data.yaml to skip reconversion when nothing meaningful changed 🧾🔒
  • Deletes orphaned images that are no longer part of the dataset after a resplit, reducing “stale background files” that can silently affect training 🗑️
• More reliable EdgeTPU exports 🧩⚙️
  • Automatically disables end2end mode for EdgeTPU exports (and logs a warning), aligning EdgeTPU with other limited backends that don’t support required ops.
• OpenVINO INT8 export dependency handling improved 📦✅
  • Adds PyTorch 2.3 detection and adjusts nncf requirements to reduce install/export conflicts across Torch versions.
• Clearer disk space error messages 💾🔎
  • Fixes incorrect “GB” reporting and now shows MB for sizes under 1GB, making download/asset errors easier to understand.

🎯 Purpose & Impact

• Faster dataset iteration for NDJSON workflows ⚡
  Resplitting a dataset no longer forces full re-downloads—saving time, bandwidth, and friction when you’re refining splits during curation.
• Fewer training gotchas from stale files 🧼
  Clearing old labels and removing orphaned images helps prevent subtle mismatches (images/labels out of sync) that can degrade training quality.
• More dependable deployment exports 📤
  EdgeTPU exports should fail less often due to unsupported end2end behavior, and OpenVINO INT8 export setups are smoother across PyTorch versions.
• Better user experience when storage is low 🧰
  Disk space errors now communicate real numbers, reducing confusion during downloads and conversions.

What's Changed

• Fix incorrect disk space in error message by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/23727
• Pin nncf<3 for PyTorch 2.2 and below by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/23726
• Disable end2end for EdgeTPU by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/23724
• ultralytics 8.4.17 NDJSON dataset re-split support by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/23735

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.16...v8.4.17

🌟 Summary (single-line synopsis)

Ultralytics 8.4.16 mainly improves Windows ↔ Linux/macOS .pt model portability by fixing pathlib pickle issues, with a small docs update for HEIC/HEIF image support 🧩💾📸

📊 Key Changes

• 🪟🖥️ Cross-platform .pt loading fix (PR #23725 by @glenn-jocher):
  • Updated torch_safe_load() to remap pathlib.WindowsPath ↔ pathlib.PosixPath during unpickling, preventing common load failures when models are created on one OS and loaded on another.
• 🧵 Checkpoint metadata made safer/more consistent:
  • model.pt_path is now stored as a plain string (str(weight)), not a Path object (applies in general checkpoint loading and YOLO NAS model wrapper code).
• 🔢 Version bump: 8.4.15 → 8.4.16.
• 📸 Docs + format clarity (PR #23717):
  • Docs now reflect that HEIC/HEIF uses pi-heif (auto-installed on first use) and AVIF is supported natively by Pillow.
  • .heif is explicitly listed among supported image formats.

🎯 Purpose & Impact

• ✅ Fewer “can’t unpickle WindowsPath/PosixPath” errors when sharing trained .pt weights between teammates or CI systems running different operating systems.
• 🔁 More reliable export and tooling workflows that depend on model.pt_path (strings behave consistently across platforms and serialization).
• 👥 Better collaboration & reproducibility for teams training on Linux servers and deploying/testing on Windows machines (or vice versa).
• 📸 Smoother HEIC/HEIF user experience thanks to clearer docs and explicit .heif support visibility (less dependency confusion).

What's Changed

• Update HEIC/HEIF dependency info and add .heif format to docs by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/23717
• ultralytics 8.4.16 Windows<>Linux model compatibility fix by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/23725

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.15...v8.4.16

🌟 Summary (single-line synopsis)

Ultralytics v8.4.15 improves real-world image ingestion (adds HEIC/HEIF + HEIF support with lazy decoding + better EXIF handling) and includes several quality fixes for metrics, exports, and docs 🖼️⚙️✅

📊 Key Changes

• HEIC/HEIF image support (priority change) 🖼️📱
  • Added ultralytics.utils.patches.image_open and monkey-patched PIL.Image.open to lazily enable HEIC/HEIF decoding via pi-heif only when needed (i.e., on first decode failure) ⏱️
  • Expanded supported extensions to include .heif (in addition to .heic) ✅
  • Removed HEIC-specific loading code from dataset loaders (now unified through the standard image reading pathway) 🧹
• Better plotting for non‑Latin text 🈶🖋️
  • Improves matplotlib defaults by prioritizing Arial Unicode MS and auto-registering .ttf fonts from the Ultralytics user config directory for stronger multilingual rendering.
• EXIF auto-rotation fix for list-of-path sources 🔄🖼️
  • Applies ImageOps.exif_transpose() when loading images from lists of paths/URLs, reducing “sideways phone photo” issues.
• Validation/metrics correctness fix (important for accuracy reporting) 🎯📈
  • Fixes SciPy-based matching (match_predictions) to maximize IoU during assignment (previously could choose the worst matches), which can affect TP/FP classification and mAP when use_scipy=True.
• IMX (Sony IMX500) export improvements 📦🤖
  • Enables IMX export on Linux ARM64 (removes an ARM64 block), adds dependency support, and improves test gating.
  • Adds segmentation as a supported IMX export/benchmark task ✅
• Usability improvement: Results.save() 💾📁
  • Adds create_dir parameter (default True) to control auto-creation of parent directories when saving annotated outputs.
• Dev/CI reliability updates 🧪🏗️
  • CI now tests newer PyTorch versions (2.8/2.9) and makes export tests less flaky by skipping TFLite export tests unless torch>=1.13 on Linux.

🎯 Purpose & Impact

• More “it just works” image input 🖼️✅
  • HEIC/HEIF (common on iPhones) can be used directly for training/inference with fewer manual conversions, while lazy loading avoids paying the dependency cost unless you actually need it.
• Fewer confusing orientation bugs 📱➡️🖥️
  • EXIF transpose handling reduces rotated/sideways images that can break visualization, training data review, and inference outputs.
• More trustworthy evaluation metrics 📈🧠
  • The IoU matching fix ensures SciPy-based evaluation behaves as intended, preventing artificially worse results or incorrect TP/FP assignment when that path is enabled.
• Better edge/export support on ARM 🧩🚀
  • IMX export on ARM64 + segmentation support expands deployment options for embedded and edge workflows.
• Smoother workflows & sharing outputs 💾✨
  • Results.save(create_dir=True) makes saving predictions into new folders simpler (fewer “directory not found” errors).
• Improved global documentation and plots 🈶📚
  • Better multilingual font defaults help teams with non‑Latin labels produce clearer charts/figures.

If you’re deploying/training with mixed photo sources (phones, HEIC libraries, multilingual labels), 8.4.15 is a practical upgrade 🛠️🚀

What's Changed

• Add guide on customizing the YOLO trainer by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/23579
• Add YOLO26 vs YOLO11 performance charts for Raspberry Pi 5 by @lakshanthad in https://github.com/ultralytics/ultralytics/pull/23625
• test: 🧪 Add TFLite export requirement for Torch version 1.13 by @onuralpszr in https://github.com/ultralytics/ultralytics/pull/23618
• Fix typos and grammar errors in documentation by @him2him2 in https://github.com/ultralytics/ultralytics/pull/23617
• feat: improve Results.save() with pathlib and optional directory creation by @ShuaiLYU in https://github.com/ultralytics/ultralytics/pull/23592
• fix: 🐞 add ExecuTorch version during model export into logging by @onuralpszr in https://github.com/ultralytics/ultralytics/pull/23620
• ci: 👷 move Conda Builds CI to monitor from conda-forge/ultralytics-feedstock to ultralytics repo by @onuralpszr in https://github.com/ultralytics/ultralytics/pull/23600
• Add https://youtu.be/GAFlmuk0fZI to docs by @RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/23624
• Apply EXIF transpose when list of paths is used as source by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/23631
• Fix scipy matching in match_predictions to maximize IoU by @Mr-Neutr0n in https://github.com/ultralytics/ultralytics/pull/23634
• Enable IMX export for ARM64 by @lakshanthad in https://github.com/ultralytics/ultralytics/pull/23636
• Enable segmentation models in IMX benchmarks by @lakshanthad in https://github.com/ultralytics/ultralytics/pull/23635
• Fix: get the correct package name for metadata.distribution by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/23641
• docs: 📝 update hyperlink for hyperparameter tuning in custom trainer guide by @onuralpszr in https://github.com/ultralytics/ultralytics/pull/23644
• Fix broken YOLO26 documentation link in vision-eye guide by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/23646
• Fix docstrings by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/23648
• docs: 📝 fix conda badge link in CI documentation by @onuralpszr in https://github.com/ultralytics/ultralytics/pull/23661
• docs: 📝 remove unneeded tokenizer_ve documentation file by @onuralpszr in https://github.com/ultralytics/ultralytics/pull/23662
• Delete orphan markdown files in build_reference.py by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/23666
• docs: 📝 remove non-existed dataset generation section from COCO12-Formats documentation by @onuralpszr in https://github.com/ultralytics/ultralytics/pull/23667
• Simplify CustomSaveTrainer example in custom trainer guide by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/23668
• Add custom trainer guide cross-reference in train.md by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/23649
• Add PyTorch 2.8 and 2.9 tests to slow CI by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/23683
• Update platform docs by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/23688
• Update Platform docs by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/23692
• Added max concurrency 16 to link checker workflow for the HTTP/2 error by @banica-ultralytics in https://github.com/ultralytics/ultralytics/pull/23690
• Add https://youtu.be/pJLXmhyuHzA to docs by @RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/23658
• Platform Docs screenshots by @sergiuwaxmann in https://github.com/ultralytics/ultralytics/pull/23697
• Add https://youtu.be/cxU5E2SkivU to docs by @RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/23701
• Update platform API docs by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/23704
• Platform Rate Limit docs by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/23706
• Fix uv conflicts between export and solutions extras by @fcakyon in https://github.com/ultralytics/ultralytics/pull/23686
• Add YOLOE agnostic_nms default behavior note by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/23707
• Document CUDA OOM auto-retry in batch-size settings by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/23708
• Document Results.save() automatic directory creation by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/23709
• Banica ultralytics http2 changes undone by @banica-ultralytics in https://github.com/ultralytics/ultralytics/pull/23712
• ultralytics 8.4.15 HEIC/HEIF image support by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/23714

New Contributors

• @him2him2 made their first contribution in https://github.com/ultralytics/ultralytics/pull/23617
• @banica-ultralytics made their first contribution in https://github.com/ultralytics/ultralytics/pull/23690

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.14...v8.4.15

🌟 Summary (single-line synopsis)

Ultralytics 8.4.14 adds Ultralytics Platform “Cancel training” support so you can stop runs quickly and still keep/upload partial results ⛔️📤

📊 Key Changes

• Platform-driven training cancellation (priority change) 🛑
  • Detects cancellation even before training starts (during session registration) and prevents wasted startup time.
  • Checks for cancellation at epoch end via a send-and-check response flow, then sets trainer.stop=True when cancelled=true.
  • Tracks cancellation state with trainer._platform_cancelled and logs clear messages.
  • Continues to upload partial artifacts/metadata when a run is cancelled (so you don’t lose everything).
• Faster responsiveness to stop requests between batches ⚡
  • Training loop now breaks if self.stop is set, allowing external stops (like Platform cancellation) to take effect sooner.
• YOLO26 Pose training stability fix 🧍‍♂️📈
  • Clamps negative rle_loss to zero to prevent loss from going negative and destabilizing training (helps avoid mAP dropping across epochs).
• Segmentation → bounding box conversion edge-case fix 🖼️📦
  • segment2box() now excludes points exactly on the image border after clipping, avoiding boxes incorrectly snapping to edges (better box regression + mask quality).
• Hyperparameter tuning crash fix 🧪🔧
  • Prevents TypeError when fitness is present but None during tuning (safer/cleaner tuning runs).
• Docs clarity: Params/FLOPs reporting after model.fuse() 📚
  • Adds notes explaining why Params/FLOPs in docs may differ from what you see locally (fused inference vs full training architecture).

🎯 Purpose & Impact

• Better control on Ultralytics Platform 🎛️
  • If you hit Cancel in the Platform, training should stop reliably and promptly—without waiting for a long delay.
• Lower wasted compute + cost 💸
  • Cancelled jobs exit earlier and more predictably, especially helpful for long training runs.
• Keeps useful outputs even when cancelling 📦
  • Partial uploads mean you can still inspect progress, logs, and intermediate artifacts after stopping.
• More stable YOLO26 Pose training ✅
  • Reduces risk of training “going backwards” due to negative loss behavior; improves consistency of pose metrics over epochs.
• Higher-quality results for segmentation-derived boxes 🎯
  • Fewer edge-snapped boxes after augmentations can improve both training quality and final predictions.
• Fewer interruptions in automated workflows 🤖
  • Tuning and reporting are more robust, reducing flaky failures in pipelines.

If you train via the Ultralytics Platform, this release is especially impactful due to the new cancellation behavior 🛑.

What's Changed

• Fix TypeError when fitness is None during hyperparameter tuning by @Mr-Neutr0n in https://github.com/ultralytics/ultralytics/pull/23603
• Clamp negative rle_loss to zero in Pose26 training by @Mr-Neutr0n in https://github.com/ultralytics/ultralytics/pull/23604
• Fix segment2box boundary condition for edge-snapped segment points by @Mr-Neutr0n in https://github.com/ultralytics/ultralytics/pull/23602
• Add footnote clarifying param/FLOP counts after model.fuse() by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/23601
• ultralytics 8.4.14 Platform training cancel feature by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/23614

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.13...v8.4.14

🌟 Summary (single-line synopsis)

Ultralytics v8.4.13 makes training more resilient by automatically recovering from CUDA out-of-memory (OOM) errors during the first epoch by retrying with a smaller batch size 🔁🧠🔥

📊 Key Changes

• Auto-retry on CUDA OOM during training (major change) 🔥🛠️
  • If a CUDA OOM happens in the first epoch on single-GPU, Ultralytics will retry up to 3 times, halving the batch size each time (down to 1).
  • Training pipeline is rebuilt after batch reduction (dataloaders + optimizer + scheduler) to continue cleanly.
• New internal training helper 🧩
  • Adds a _build_train_pipeline() method to rebuild loaders/optimizer/scheduler when batch size changes (used by the new OOM recovery flow).
• More reliable ONNX export for OBB + NMS 📦✅
  • When exporting OBB (oriented bounding boxes) to ONNX with NMS enabled, simplify=True is now forced to avoid a known runtime issue (TopK-related error in some ONNX Runtime versions).
• DGX system detection + TensorRT handling 🖥️⚙️
  • Adds is_dgx() detection and uses it (along with Jetson JetPack 7) to trigger a TensorRT version check/reinstall path for better export reliability on those systems.
• Packaging stability fix: pin setuptools 🧰🔒
  • Pins build requirements to setuptools<=81.0.0 to avoid breakages introduced by newer setuptools versions (notably affecting tensorflow.js export tooling).
• Docs & examples refresh (YOLO26 messaging + tracking content) 📚🎥
  • Tracking docs now embed a newer multi-object tracking video featuring YOLO26 + BoT-SORT/ByteTrack.
  • Exporter docs/examples updated to show YOLO26 (yolo26n.pt) and mention ExecuTorch/Axelera export options (documentation signposting).
• Example dependency update 🔄
  • Updates protobuf in the RT-DETR ONNX Runtime Python example.

🎯 Purpose & Impact

• Fewer training crashes for everyday users 🙌🔥
  • If you start training with a batch size that’s slightly too large for your GPU, Ultralytics can now self-correct and continue instead of failing immediately—especially helpful for beginners and for “first-epoch spikes” in memory use.
• Less manual trial-and-error 🎯
  • Reduces the common loop of “OOM → lower batch → restart training,” saving time and frustration.
• More dependable deployment exports 🚀
  • ONNX exports for OBB models with embedded NMS should work more reliably out of the box, with fewer runtime surprises.
• More predictable builds/CI 🧱
  • Pinning setuptools helps prevent sudden packaging/tooling failures across environments.
• Clearer guidance aligned with YOLO26 🧭
  • Docs and examples increasingly steer users toward YOLO26 as the recommended model for training, tracking, and export workflows.

What's Changed

• feat: 🚀 NVIDIA DGX device variants check by @onuralpszr in https://github.com/ultralytics/ultralytics/pull/23573
• Add https://youtu.be/qQkzKISt5GE to docs by @RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/23582
• Bump protobuf from 6.31.1 to 6.33.5 in /examples/RTDETR-ONNXRuntime-Python in the pip group across 1 directory by @dependabot[bot] in https://github.com/ultralytics/ultralytics/pull/23572
• docs: 📝 exporter documentation for new model formats and examples updated by @onuralpszr in https://github.com/ultralytics/ultralytics/pull/23585
• Force simplify=True for OBB export with NMS by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/23580
• Pin setuptools version by @Burhan-Q in https://github.com/ultralytics/ultralytics/pull/23589
• ultralytics 8.4.13 Retry smaller batch on training CUDA OOM by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/23590

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.12...v8.4.13

🌟 Summary (single-line synopsis)

Ultralytics v8.4.12 streamlines YOLOE-26 class/text-prompt handling to avoid redundant updates, while improving multispectral (grayscale) training reliability and multi-GPU auto-selection 🧠⚡️🖥️

📊 Key Changes (major updates)

• YOLOE-26: Skip redundant set_classes() when prompts already match 🧠⚡️
  • Adds a guard in ultralytics/models/yolo/model.py to avoid re-running set_classes() if model.names already matches the requested classes.
  • Only regenerates text embeddings (get_text_pe(classes)) when a real class change is needed.
  • Ensures the predictor stays in sync by setting self.predictor.model.names = self.model.names.
• More robust GPU auto-selection under multi-process starts 🎲🖥️
  • Adds a random tie-break when GPUs look equally idle, reducing the chance that multiple jobs pick the same GPU at launch (ultralytics/utils/autodevice.py).
• Fix disk caching for grayscale/multispectral images 💾🖼️
  • Disk cache now respects the dataset’s OpenCV read mode (flags=self.cv2_flag), preventing cached .npy files from silently changing input format (ultralytics/data/base.py).
• Fix AutoBatch for non-RGB inputs (grayscale/multispectral) 🧪⚙️
  • AutoBatch profiling now uses the model’s configured input channels (model.yaml["channels"]) instead of assuming 3-channel RGB (ultralytics/utils/autobatch.py).
• Tests: better grayscale coverage + cleanup 🧪🧹
  • Adds CUDA training coverage for coco8-grayscale.yaml, uses cache="disk" in grayscale tests, and cleans up cached .npy files to prevent flaky test interactions.
• Docs/examples updated to YOLO26 naming 📝🚀
  • Exporter and error-message examples now reference YOLO26 (recommended) instead of YOLO11 (e.g., yolo26n.onnx, yolo26n_openvino_model, etc.).
• Confusion matrix plot fixes for classification 📉✅
  • Removes unintended “background” labeling behavior for classification confusion matrices and makes tick labeling more robust for large class counts (ultralytics/utils/metrics.py).
• Minor tracker cleanup 🧩
  • Simplifies deque initialization in BoT-SORT tracker (ultralytics/trackers/bot_sort.py).
• Docker dependency tweak 🐳
  • Swaps wandb for nvidia-ml-py in the Dockerfile to better support NVIDIA GPU querying/metrics in container builds.

🎯 Purpose & Impact (why it matters)

• Faster, more stable YOLOE-26 inference loops 🚀
  If you repeatedly run predictions with the same text prompts/classes, v8.4.12 avoids unnecessary reconfiguration and embedding regeneration—improving performance and reducing potential side effects.
• More reliable multi-job GPU usage 🖥️
  When launching multiple trainings/inference processes at once, GPU selection is less likely to “pile onto” the same device due to tie conditions.
• Better support for grayscale/multispectral workflows 🖤🤍
  Training with cache="disk" and using AutoBatch on non-3-channel inputs is now more correct and consistent—reducing mismatches, mis-profiling, and possible crashes.
• Clearer docs aligned with the recommended model 📘
  Updated examples reduce copy/paste confusion by pointing users to YOLO26, the latest stable and recommended Ultralytics model family.

What's Changed

• Round-robin GPU autoselect by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/23544
• Fix disk caching with multispectral images by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/23545
• Fix AutoBatch with multispectral images by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/23546
• Remove cached *.npy files in grayscale tests by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/23554
• docs: 📝 update exporter file yolo11 terms to newest yolo26 by @onuralpszr in https://github.com/ultralytics/ultralytics/pull/23556
• Remove "background" class from classification confusion matrix by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/23547
• ultralytics 8.4.12 YOLOE-26 skip set_classes if text prompts already set by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/23552

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.11...v8.4.12

🌟 Summary (single-line synopsis)

Ultralytics v8.4.11 makes Ultralytics Platform training uploads far more reliable by adding retryable model uploads + metrics posting, plus a new reusable upload helper and docs 🚀📶

📊 Key Changes

• Robust Platform upload + metrics retries (PR #23538 by @glenn-jocher) 🔁
  • Added a new utility: ultralytics/utils/uploads.py with safe_upload() that supports retries, timeouts, and optional progress bars ⏳📦
  • Platform callback networking hardened:
    • _send() now retries transient failures, uses a longer timeout (30s), and does not retry most 4xx client errors (except common transient ones like timeout/rate-limit) 🛡️
    • _upload_model() now retries signed-URL retrieval and uses safe_upload() for the actual upload to cloud storage ☁️
  • Platform now receives only the remote cloud path (gcsPath) as modelPath (instead of accidentally sending a local filesystem path) ✅
  • Training end now uploads the best model with a progress bar enabled by default 📊
  • Added a new reference doc page for uploads utilities and linked it into the docs nav 📚
• Windows OpenVINO export test stability fix (PR #23528 by @lakshanthad) 🪟✅
  • Ensures OpenVINO export tests use unique output filenames on Windows to prevent file locking/collision issues in CI 🧪

🎯 Purpose & Impact

• Fewer failed Platform syncs during training due to flaky networks or transient API/upload hiccups 📶➡️✅
  Impact: your experiment tracking, logs, metrics, and model artifacts are more likely to show up correctly on the Ultralytics Platform.
• More reliable large checkpoint uploads (with retries + progress feedback) 📦⏫
  Impact: less manual re-running/re-uploading when training finishes, especially on slower or unstable connections.
• Cleaner, safer metadata sent to Platform 🧹
  Impact: Platform gets a downloadable cloud path, not a meaningless local path like C:\... or /home/....
• Improved CI stability on Windows for OpenVINO exports 🪟
  Impact: no change for most end users, but faster/cleaner releases and fewer flaky test failures.

What's Changed

• Fix OpenVINO failures on Windows fast CI tests by @lakshanthad in https://github.com/ultralytics/ultralytics/pull/23528
• ultralytics 8.4.11 Platform model uploads retries by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/23538

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.10...v8.4.11

🌟 Summary (single-line synopsis)

Ultralytics 8.4.10 improves YOLOE-26 out-of-the-box predictions by defaulting to class-agnostic NMS, plus several export, OpenVINO, HUB, SAM, and docs reliability upgrades 🚀🧠

📊 Key Changes

• 🧠 YOLOE-26 inference default change (most important): YOLOE predictions now default to agnostic_nms=True, reducing overlapping duplicate boxes across different classes (set in ultralytics/models/yolo/model.py).
• 📦 Version bump: 8.4.9 → 8.4.10.
• 🍏 CoreML export feature: CoreML export now correctly honors agnostic_nms (previously the argument existed but wasn’t applied in the CoreML NMS pipeline).
• ⚙️ TensorRT setup reliability: added check_tensorrt() helper and used it in export + tests to reduce missing-dependency failures; also adds special handling to pin compatible TensorRT on Jetson JetPack 7 / CUDA 13 ARM for RT-DETR exports.
• 🏎️ OpenVINO bugfix: fixes an UnboundLocalError risk by making the inference-mode check safe when dynamic is false.
• ☁️ Ultralytics HUB robustness: HUB session now raises clearer HUBModelError (instead of generic ValueError or silently returning None) when loading/creating HUB models fails.
• 🧩 SAM results metadata fix: improves how class names are generated (especially for single-object outputs and “visual” naming mode).
• 🛡️ Instances safety + debugging: avoids crashes when segments=None during zero-area filtering, and adds a helpful Instances.__repr__() for easier debugging.
• 📚 Docs improvements: YOLOE examples simplified to model.set_classes(["person", "bus"]) (no manual text embedding handling), ExecuTorch docs updated with YOLO26 benchmarks, and Benchmark docs video link refreshed.

🎯 Purpose & Impact

• 🎯 More consistent YOLOE predictions by default: class-agnostic NMS helps suppress duplicate overlapping detections across classes, improving “it just works” behavior for new users.
• ⚠️ Behavior change to expect: YOLOE outputs may differ vs. 8.4.9 (some boxes previously kept may now be suppressed). If you relied on per-class duplicates, explicitly set agnostic_nms=False.
• 🍏 More predictable exports across formats: CoreML now matches other exporters when agnostic_nms is enabled—fewer surprises when deploying to Apple devices.
• 🤖 Smoother Jetson/TensorRT experience: fewer export failures and less manual dependency wrangling, especially on newer Jetson stacks (JetPack 7).
• 🧯 Fewer runtime errors: OpenVINO, Instances filtering, HUB model handling, and SAM naming fixes reduce “random crash” scenarios and make failures easier to diagnose.
• 🧼 Cleaner onboarding + examples: YOLOE prompting is easier to copy/paste, and YOLO26 guidance in docs stays aligned with the recommended models.

You can update with: pip install -U ultralytics ⬆️✅

What's Changed

• fix: 🐞 Add requirement check for TensorRT in test_export_engine_matrix by @onuralpszr in https://github.com/ultralytics/ultralytics/pull/23496
• Add https://youtu.be/UF7pYdLSMng to docs by @RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/23497
• Fix autobackend.py when using OpenVINO by @zboszor in https://github.com/ultralytics/ultralytics/pull/23505
• Force update tensorrt on CUDA 13 ARM to 10.15.x to fix a bug with RT-DETR exports by @lakshanthad in https://github.com/ultralytics/ultralytics/pull/23499
• Update executorch doc with YOLO26 benchmarks on Raspberry Pi 5 by @lakshanthad in https://github.com/ultralytics/ultralytics/pull/23500
• 🐞 fix remove_zero_area_boxes instance when segment is None by @Bovey0809 in https://github.com/ultralytics/ultralytics/pull/23495
• Add repr for Instance Class for devlopers who needs better debugging … by @Bovey0809 in https://github.com/ultralytics/ultralytics/pull/23473
• Update YOLOE docs example by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/23526
• Remove debug asserts from TVP loss classes by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/23523
• Improve HUB session error handling with HUBModelError by @ahmet-f-gumustas in https://github.com/ultralytics/ultralytics/pull/23512
• Fix names generation for single object detection in SAM3SemanticPredictor by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/23513
• CoreML Export: Add Support for Class-Agnostic NMS by @K-saif in https://github.com/ultralytics/ultralytics/pull/23493
• ultralytics 8.4.10 YOLOE-26 default agnostic_nms=True by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/23525

New Contributors

• @K-saif made their first contribution in https://github.com/ultralytics/ultralytics/pull/23493
• @zboszor made their first contribution in https://github.com/ultralytics/ultralytics/pull/23505

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.9...v8.4.10

🌟 Summary (single-line synopsis)

Ultralytics v8.4.9 improves segmentation CopyPaste augmentation reliability (especially for grayscale/hyperspectral-style single-channel inputs) and strengthens the broader export/training ecosystem with better ExecuTorch + Torch support 🚀🧩

📊 Key Changes

• 🧪 (Priority) Segmentation CopyPaste mask fix in ultralytics/data/augment.py:
  • Builds the contour mask as a 2D (H×W) array instead of matching the full image tensor shape.
  • Draws contours onto a single-channel mask using value 1 (instead of (1,1,1)), avoiding channel mismatches.
• 🧾 Version bump: 8.4.8 → 8.4.9 🔢
• 📦 NDJSON datasets always re-download when using ul://... URIs (instead of silently reusing a cached .ndjson) 🧹⬇️
• ⚙️ ExecuTorch support improvements:
  • Removes the <2.10 PyTorch upper bound in dependencies (Windows still excludes torch==2.4.0) 🔓
  • Centralizes ExecuTorch install checks via check_executorch_requirements() to reduce platform-specific install failures 🧩
• ⚡ Training + compile coverage: CUDA test now trains with compile=True, and stride calculation is simplified for more predictable behavior 🧠🔥
• 📚 Docs refresh: Export docs video updated to a more relevant YOLO26 export/deployment tutorial 🎥

🎯 Purpose & Impact

• 🧩 More robust segmentation augmentation: prevents mask shape/channel bugs that can break or silently degrade CopyPaste-based training—especially when images are single-channel (common in grayscale and some hyperspectral pipelines) 📷✅
• 📈 More reliable segmentation training: fewer augmentation edge cases means smoother training runs and more consistent results 🏋️‍♂️📊
• 🧹 Less “stale dataset” confusion: always refreshing .ndjson metadata helps ensure you train/evaluate on the latest Ultralytics-hosted dataset state when using ul://... 🔄🧪
• 🚀 Better export/inference readiness on modern stacks: improved ExecuTorch requirement handling + Torch 2.10+ compatibility reduces install friction and improves deployment workflows 📦⚙️
• ⚡ Safer compile=True usage: better test coverage helps catch regressions for users relying on PyTorch compile for potential speedups 🧠🛠️

What's Changed

• Always redownload NDJSON datasets by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/23458
• feat(executorch): Add support for Torch 2.10 in ExecuTorch requirements and imports and remove torch<2.10 pin by @onuralpszr in https://github.com/ultralytics/ultralytics/pull/23344
• Add missing type hints to checks.py and dist.py by @ahmet-f-gumustas in https://github.com/ultralytics/ultralytics/pull/23485
• Add https://youtu.be/KGHYU-MKYeE to docs by @RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/23463
• Fix training with compile by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/23479
• refactor: ♻️ Simplify requirement checks for executorch by @onuralpszr in https://github.com/ultralytics/ultralytics/pull/23475
• ultralytics 8.4.9 Hyperspectral CopyPaste augmentation by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/23471

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.8...v8.4.9

🌟 Summary (single-line synopsis)

Ultralytics v8.4.8 makes end-to-end (NMS-free) YOLO26/YOLOv10 inference actually honor max_det and agnostic_nms, giving you predictable control over how many detections you get and how classes are handled ✅🚀

📊 Key Changes

• End2end now supports max_det + agnostic_nms (PR #23396 by @Y-T-G) 🎛️
  • Adds a safe set_head_attr(**kwargs) helper to set head/last-layer attributes like end2end, max_det, agnostic_nms 🧩
  • When end2end=True, predict and val now push max_det + agnostic_nms into the model head, so these flags actually take effect ⚙️
  • Adds agnostic_nms support inside the Detect head (affects top-k selection logic used in end-to-end mode) 🎯
  • Export pipeline also carries agnostic_nms into the exported head for better parity between Python and exported models 📦
• Clearer end2end controls across CLI/docs 📚
  • Adds end2end to config/defaults and docs for predict/val/export args 🧾
  • Expands tests to run export/predict/val with end2end=True/False ✅🧪
• Notable additional fixes included in this tag (smaller but impactful) 🛠️
  • Export reliability: exporter auto-disables end2end for formats that don’t support required ops (RKNN/NCNN/ExecuTorch/Paddle/IMX) and stores end2end in metadata 🧰
  • YOLOE/YOLO-World training usability: allows passing a YAML file (or Path) for multi-source data configs 🗂️
  • Docs improvements: new Compare section + cleaner navigation; sitemap auto-fills missing pages 🧭
  • Stability fixes: rotated-OBB small-box robustness, DDP unwrap fixes for pose validation, PyTorch 1.9 compatibility adjustment, OpenVINO/ONNX example bugfixes 🔧

🎯 Purpose & Impact

• More predictable end-to-end inference results 🧠
  • max_det now reliably limits outputs in end2end mode (no more “why didn’t it cap?” confusion).
• Better class-agnostic behavior when you need it 🧹
  • agnostic_nms becomes meaningful in end2end workflows, helpful when classes overlap heavily or you want “best boxes regardless of class”.
• Closer parity between Python and exported deployments 🚚
  • Exported models are less likely to behave differently from local inference due to missing head settings.
• Fewer surprises across formats & environments ✅
  • Safer exporting to constrained runtimes, more robust training/validation in edge cases, and improved docs discoverability for new users.

If you’re using YOLO26 (the recommended model family) in end-to-end mode, this release is a practical quality-of-life upgrade. 🚀🙂

What's Changed

• Update optimizer selection strategy by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/23350
• Fix broken links in NVIDIA Jetson documentation by @lakshanthad in https://github.com/ultralytics/ultralytics/pull/23362
• Add ability to override end2end mode by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/23333
• Update small obb candidates selection by @lmycross in https://github.com/ultralytics/ultralytics/pull/23378
• Update coco12-formats predict tests by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/23374
• fix openvino sample for coordinates by @jules-ai in https://github.com/ultralytics/ultralytics/pull/23366
• Fix index issue of strides in select_candidates_in_gts by @PT0X0E in https://github.com/ultralytics/ultralytics/pull/23369
• Add note about SAM3 tokenizer error due to incorrect clip package by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/23368
• Fix stride_val type issue in STAL by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/23390
• Export non-e2e version by default for rknn/executorch/paddle/imx formats by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/23312
• fix: add unwrap_model to fix DDP AttributeError by @lmycross in https://github.com/ultralytics/ultralytics/pull/23392
• Fix: use torch.no_grad for torch==1.9.0 to resolve training issue by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/23395
• Simplify Docs nav by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/23398
• Fix swapped input_height and input_width in ONNX example - Fixes #23126 by @ahmet-f-gumustas in https://github.com/ultralytics/ultralytics/pull/23402
• Add models compare pages to Docs by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/23406
• Fix visual prompt training for YOLOE26 by @ShuaiLYU in https://github.com/ultralytics/ultralytics/pull/23401
• Add https://youtu.be/7lZa3Yi2kbo to docs by @RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/23388
• Classify train fixes by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/23420
• Fix/yoloe text model attribute in yoloe trainer from scratch by @ShuaiLYU in https://github.com/ultralytics/ultralytics/pull/23428
• Fix names generation in SAM3VideoSemanticPredictor by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/23434
• Update NVIDIA Jetson Thor benchmarks with YOLO26 by @lakshanthad in https://github.com/ultralytics/ultralytics/pull/23443
• Save close_mosaic as integer in best_hyperparameters.yaml by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/23435
• Feat: support multiple data config via yaml for YOLOE training by @ShuaiLYU in https://github.com/ultralytics/ultralytics/pull/23427
• Fix YOLOE training test by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/23448
• Fix SMTP port parameter syntax in SecurityAlarm by @ahmet-f-gumustas in https://github.com/ultralytics/ultralytics/pull/23453
• Disable shuffle only if batch shapes mismatch with rect=True by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/23419
• ultralytics 8.4.8 Support max_det and agnostic_nms for end2end by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/23396

New Contributors

• @PT0X0E made their first contribution in https://github.com/ultralytics/ultralytics/pull/23369

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.7...v8.4.8

🌟 Summary

Ultralytics v8.4.7 adds AVIF training support and a new COCO12-Formats mini-dataset to continuously verify that all supported image types load correctly end-to-end (especially in CI) 🧪🖼️

📊 Key Changes

• AVIF is now a supported training image format ✅📷
  • IMG_FORMATS expanded to include avif
• More robust image decoding for modern formats 🛡️
  • Adds a Pillow-based fallback image reader (_imread_pil) for cases where OpenCV can’t decode AVIF/HEIC
• New “COCO12-Formats” dataset + generator script 🧰
  • A tiny 12-image dataset covering all supported extensions (AVIF, BMP, DNG, HEIC, JP2, JPEG, JPG, MPO, PNG, TIF, TIFF, WebP)
  • Includes a dataset YAML and a generator (generate_coco12_formats.py) to build the set for testing/validation
  • Adds a dedicated docs page and links it in the dataset index 📚
• Training and tooling reliability improvements 🔧
  • Fixes loss image-size calculation for OBB and Pose26 multi-scale training 📐
  • Fixes semantic segmentation masking/indexing for overlapping instances 🧩
  • More accurate optimizer parameter-group logging 🧾
  • Adds retry + validation for dataset image downloads (NDSON) 🔁📥
• Install/export stability & experiment tracking 📦📈
  • Pins torch < 2.10 to prevent ExecuTorch export incompatibilities 🧷
  • Saves runs under runs/ by default for relative projects + improves W&B traceability 🗂️
• Docs updates (large) 📝⚡
  • Docs images migrated to a CDN and improved alt text for accessibility/SEO

🎯 Purpose & Impact

• Train directly on AVIF images without manual conversion 🎉
  • Useful for modern web/phone pipelines where AVIF is common
• Fewer “it works on my machine” image-loading failures 🧯
  • The new COCO12-Formats dataset is designed to catch format-specific decoding issues early, especially in CI/CD
• Better resilience across environments 🌍
  • Pillow fallback reduces dependence on how OpenCV was built (many prebuilt OpenCV wheels lack AVIF support)
• More reliable training & clearer debugging 🔍
  • Fixes reduce subtle scaling bugs (Pose/OBB), segmentation masking issues, and misleading optimizer logs
• Smoother installs/exports and cleaner run management 🚀
  • Dependency pin helps prevent breakage; run folder changes improve organization and experiment tracking

What's Changed

• Fix docs images content disposition by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/23322
• Docs images cleanup by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/23323
• Fix the length of optimizer params group print by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/23314
• Platform docs by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/23338
• Route outputs to runs/ by default and improve W&B run traceability by @artest08 in https://github.com/ultralytics/ultralytics/pull/23026
• fix image size calculation in OBB and Pose26 loss for multi-scale training by @lmycross in https://github.com/ultralytics/ultralytics/pull/23342
• fix: 🐞 pin torch version to <2.10 for executorch export by @onuralpszr in https://github.com/ultralytics/ultralytics/pull/23353
• fix semantic segmentation loss indexing error by @lmycross in https://github.com/ultralytics/ultralytics/pull/23330
• Retry NDSON download failures by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/23354
• Fix multi_scale range sampling and clarify docs by @artest08 in https://github.com/ultralytics/ultralytics/pull/23284
• ultralytics 8.4.7 AVIF training and new COCO12-Formats dataset by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/23358

New Contributors

• @artest08 made their first contribution in https://github.com/ultralytics/ultralytics/pull/23026

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.6...v8.4.7

🌟 Summary (single-line synopsis)

Ultralytics v8.4.6 is a reliability-focused release that fixes a multi-GPU DDP training crash and improves Ultralytics HUB/Platform dataset URI handling, plus several documentation upgrades 📈🛠️

📊 Key Changes

• ✅ Fixed DDP multi-GPU training crash (PR #23301 by @pfabreu)
  • Added missing PosixPath import to the generated temporary DDP training script (ultralytics/utils/dist.py).
  • Prevents NameError: name 'PosixPath' is not defined when model paths are passed as PosixPath objects (common on Linux) during Distributed Data Parallel training 🧩🧯
• ⏱️ More robust dataset URI resolution for large Ultralytics HUB/Platform datasets (PR #23305 by @glenn-jocher)
  • ul://... dataset URLs now allow a much longer server-side preparation time (NDJSON generation) to avoid premature timeouts 🐢📦
• 🧹 Developer quality: BYTETracker typing modernized (PR #23304)
  • Switched legacy # type: comments to inline type annotations in BYTETracker (no behavior change) 🧠🔧
• 📚 Docs + Platform guidance improvements (PR #23294, #23298)
  • Added new docs for Activity Feed and Trash/Restore, refreshed billing/GPU pricing, added “Try on Ultralytics Platform” callouts for YOLO11 / YOLO26 / YOLOv5 / YOLOv8, plus small typo fixes ✍️🚀

🎯 Purpose & Impact

• DDP training is stable again on multi-GPU setups 🎛️✅
  • If you upgraded to 8.4.5 and saw failures when launching DDP training (especially with models like yolo26n.pt passed as a Path/PosixPath), this release removes that blocker—less downtime, fewer confusing crashes.
• Fewer flaky failures when using ul://... datasets from Ultralytics HUB/Platform 🌐✅
  • Large datasets can take longer to prepare; the smarter timeout reduces false “timeout” errors during dataset resolution.
• Better usability and discoverability in docs 🧭📘
  • Clearer Platform workflows (Activity Feed, Trash/Restore) and more direct “try it now” paths for major YOLO models.

For more context, see the v8.4.6 release tag on GitHub.

What's Changed

• Improve Platform docs by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/23294
• Fix typos in YOLO26 and YOLOE docs by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/23298
• Longer NDJSON generation timeout by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/23305
• Modernize type comments to inline annotations in byte_tracker.py by @ahmet-f-gumustas in https://github.com/ultralytics/ultralytics/pull/23304
• ultralytics 8.4.6 Add missing PosixPath import in DDP train file gen by @pfabreu in https://github.com/ultralytics/ultralytics/pull/23301

New Contributors

• @ahmet-f-gumustas made their first contribution in https://github.com/ultralytics/ultralytics/pull/23304

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.5...v8.4.6

🌟 Summary (single-line synopsis)

Ultralytics 8.4.5 makes 2D pose Results.summary() safer and more compatible by correctly handling keypoints with or without visibility flags 🧍‍♂️✅

📊 Key Changes

• 🧩 Pose Results summary now supports keypoints missing visibility: Results.summary() checks kpt.has_visible and only outputs "visible" when it exists (otherwise returns just "x" and "y"), preventing crashes in mixed keypoint formats.
• 🔖 Version bump: 8.4.4 → 8.4.5.
• 📚 Docs & examples shift to YOLO26: Kaggle links, Ultralytics Platform docs, and multiple notebooks now point to YOLO26 as the recommended current model family 🚀.
• 🗂️ NDJSON dataset docs improved: clearer per-task examples/tabs for Detect/Segment/Pose/OBB/Classify, including pose visibility explanations 🧾👀.
• 🧹 Dependency cleanup: removed optional hub-sdk extra from pyproject.toml (simplifies installs for users who don’t need it) 📦✂️.

🎯 Purpose & Impact

• 🛡️ Fewer runtime errors in pose pipelines: exporting/logging/serving pose outputs as JSON via Results.summary() won’t fail when datasets/models omit visibility/confidence per keypoint (common in non-COCO formats).
• 🔄 More robust downstream integrations: safer summaries improve reliability for APIs, analytics, logging, and dataset export workflows that depend on consistent serialization.
• 📈 Clearer “recommended model” guidance: shifting docs and notebooks toward YOLO26 reduces confusion for new users and aligns examples with the current default direction.
• 🧰 Better dataset onboarding: expanded NDJSON docs help users format annotations correctly across tasks, reducing trial-and-error when building datasets.

What's Changed

• Add YOLO26 Kaggle models by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/23289
• YOLO26 Google Colab notebook by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/23290
• Update YOLO26 Colab Notebook by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/23291
• Update NDJSON docs by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/23292
• ultralytics 8.4.5 2D Pose Result.summary() support by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/23293

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.4...v8.4.5

🌟 Summary (single-line synopsis)

Ultralytics v8.4.4 refines MuSGD training behavior (better scaling for short vs. long runs) and includes several quality-of-life fixes for exports, segmentation outputs, logging, and filesystem side effects 🚀🧠

📊 Key Changes

• 🧠 MuSGD optimizer scaling update (PR #23279 by @Laughing-q)
  MuSGD scale factors are now chosen more appropriately based on total training iterations:

  • > 10,000 iterations: (muon=0.1, sgd=1.0)
  • ≤ 10,000 iterations: (muon=0.5, sgd=0.5)
    This is a behavior change in ultralytics/engine/trainer.py.

• 📷 Sony IMX500 export compatibility improved (PR #23266 by @Laughing-q)

  • IMX exporter now accepts multiple valid layer counts (a set of allowed values) instead of requiring exactly one, reducing false “unsupported model” failures ✅
  • IMX500 docs/examples are refreshed to focus on YOLO11 (model names and paths updated).

• 🧩 Segmentation proto/output handling fixed across backends (PR #23241 by @Laughing-q)

  • Segmentation post-processing now pulls the proto masks from the correct output index for both PyTorch and exported models.
  • TensorFlow segmentation handling in AutoBackend now triggers based on task == "segment" (more reliable than guessing from output shape/length).

• 📁 No more empty run folders when save=False (PR #23268 by @Y-T-G)
  get_save_dir() no longer auto-creates directories while computing a unique run path—reduces unwanted “predict” folders appearing on disk 🧹

• 📈 TensorBoard OBB graph logging made safer (PR #23276 by @Y-T-G)
  TensorBoard graph logging now uses smart_inference_mode() to avoid gradient tracking—often less memory/overhead and fewer callback edge cases ⚙️

• 🔗 Docs branding/link refresh (PR #23283 by @glenn-jocher)
  Multiple repo/docs links now point to the Ultralytics Platform entry point (and wording updated accordingly) 🧭

🎯 Purpose & Impact

• 🚀 More consistent training dynamics with MuSGD
  If you use MuSGD (muon/sgd), you may see improved stability and convergence depending on whether your run is short or long—especially around the 10k-iteration threshold 🧠📉

• ✅ Fewer export surprises (IMX500 + segmentation exports)
  IMX500 exports should fail less often due to minor layer-count differences, and segmentation outputs should be more consistent across PyTorch/exported/TF backends 📦🧩

• 🧹 Cleaner local runs and tooling behavior
  Computing a save directory no longer creates folders prematurely, which is especially helpful for dry runs, scripts, and save=False predictions 📁✨

• ⚡ More robust logging during training
  TensorBoard graph logging becomes less intrusive and more reliable (particularly for OBB setups) 📊🛡️

What's Changed

• Fix YOLOv8/YOLO11 IMX export by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/23266
• Disable automatic directory creation with mkdir by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/23268
• Use YOLO26n for benchmark tests by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/23241
• Fix TensorBoard error with OBB by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/23276
• Replace HUB -> Platform links by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/23283
• ultralytics 8.4.4 MuSGD optimizer scale factor update by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/23279

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.3...v8.4.4

🌟 Summary (single-line synopsis)

Ultralytics v8.4.3 boosts Ultralytics Platform NDJSON dataset downloads/conversion speed 🚀, improves training metric correctness 🧠, and refreshes defaults/docs around YOLO26 📚.

📊 Key Changes

• 🚀 Faster NDJSON → YOLO dataset conversion (Ultralytics Platform data) — PR #23257 by @glenn-jocher
  • Lazy-loads aiohttp only when NDJSON conversion is used (faster startup, fewer unnecessary deps) 📦
  • Simplifies async image download code and improves concurrency scaling to match dataset size ⚡
  • Small robustness improvement for Platform URL handling 🌐
  • Version bump: 8.4.2 → 8.4.3 🔖
• 🌐 Configurable Platform base URL — PR #23256 by @glenn-jocher
  • Adds ULTRALYTICS_PLATFORM_URL to point callbacks/links to staging or local environments 🧪
• 🏷️ Defaults and examples move to YOLO26 — PR #23242 by @Laughing-q
  • Default YOLO()/CLI fallback model becomes yolo26n.pt and many docs/examples follow suit ✅
• ⚙️ Training optimizer and warmup logic made more reliable — PR #23234 by @Laughing-q
  • Parameter groups are explicitly labeled; warmup LR now targets the bias group by name (not by position) 🔥
  • “Auto” optimizer strategy simplified around MuSGD with improved defaults 🎛️
• 🧍 Pose training logs are more accurate — PR #23230 by @lmycross
  • Only reports rle_loss when the model actually supports it (avoids confusing metrics) 🧾
• 🧩 IMX inference/export consistency improvements — PR #23235 by @Laughing-q
  • Anchor/stride refresh is more robust for changing input shapes; IMX decode path simplified 🔧
• ✅ Fix duplicated Results.summary() entries — PR #23218 by @xusuyong
  • Prevents duplicated rows in summaries (cleaner analytics/logging) 🧹
• 📈 Benchmark tables clarified with end-to-end (e2e) metrics — PR #23238
  • README tables now clearly distinguish e2e evaluation metrics for YOLO26 tasks 📊
• 🍓 Raspberry Pi 5 guide updated with YOLO26 benchmarks — PR #23227
  • Refreshes benchmark formats and adds ExecuTorch results 📌

🎯 Purpose & Impact

• 🚀 If you use Ultralytics Platform datasets (NDJSON): faster, cleaner dataset conversion and downloads means quicker “time-to-train” and fewer download bottlenecks—especially on larger datasets.
• 📦 If you don’t use NDJSON conversion: lazy dependency loading reduces unnecessary imports and can make the package feel lighter/faster in common workflows.
• 🌐 For teams using staging/local Platform: ULTRALYTICS_PLATFORM_URL makes it much easier to test integrations without patching code.
• ⚙️ For training reliability: optimizer/warmup changes reduce edge-case misconfiguration and make training behavior more predictable across models/runs.
• 🧾 For pose users: loss reporting now matches model capability, improving trust in logs/metrics during training.
• 📚 For new users: YOLO26 becomes the “default path” in examples, reducing confusion about which model name to start with.

What's Changed

• Fix: dynamically adapt pose loss names to model type by @lmycross in https://github.com/ultralytics/ultralytics/pull/23230
• Update auto optimizer strategy and fix mismatched warmup_lr by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/23234
• Update IMX inference wrapper by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/23235
• Update README.md metrics table by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/23238
• Update Raspberry Pi 5 docs with YOLO26 benchmarks by @lakshanthad in https://github.com/ultralytics/ultralytics/pull/23227
• Use yolo26n in docs and examples by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/23242
• Remove duplicate result appending logic by @xusuyong in https://github.com/ultralytics/ultralytics/pull/23218
• Platform URL ENVs by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/23256
• ultralytics 8.4.3 Faster Platform NDJSON downloads by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/23257

New Contributors

• @xusuyong made their first contribution in https://github.com/ultralytics/ultralytics/pull/23218

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.2...v8.4.3

🌟 Summary (single-line synopsis)

Ultralytics v8.4.2 mainly fixes Ultralytics Platform (ul:// / NDJSON) classification training by converting datasets into the correct on-disk layout and validating them properly, plus a few quality-of-life and docs/CI tweaks 🛠️✅

📊 Key Changes

• (Most important) Platform Classification Training Fix (PR #23217, @glenn-jocher) 🧩🏷️
  • NDJSON conversion now detects task == "classify" and creates an ImageNet-style folder layout: {split}/{class_name}/... (instead of images/ + labels/ used for detection-style tasks).
  • Detection/segmentation/pose/OBB conversions remain unchanged (still producing data.yaml, images/{split}/, labels/{split}/).
  • More robust downloads: creates parent directories before saving images pulled from URLs.
  • Trainer update: Trainer.get_dataset() now always resolves ul:// / .ndjson into a local dataset first, then runs the correct dataset validation for the task.
• Classification results: Results.summary() now returns top-5 (PR #23215, @glenn-jocher) 🧠📋
  • For classification predictions, summary() is intended to return top-5 classes + confidences instead of only top-1.
  • Note: the diff shows each top-5 entry being appended twice (likely an unintended duplication) ⚠️
• Allow overriding save_dir reliably (PR #23191, @Y-T-G) 📁🔧
  • save_dir is now treated as an allowed override/config key, so setting output directories via CLI/Python overrides is more consistent.
• CI/tests: re-enable NCNN export tests on ARM64 (PR #23214, @lakshanthad) 🧪💻
  • Removes an ARM64 skip so NCNN export tests run on Apple Silicon/ARM servers (still guarded against PyTorch < 2.0 due to known instability).
• Docs and branding/link fixes (PR #23215) 📚🔗
  • Updates Kaggle model badge to YOLO11 (branding), fixes a billing link, and updates a cloud-training docs path.

🎯 Purpose & Impact

• Platform classification training is unblocked 🎉
  • If you train classification models from Ultralytics Platform datasets (ul://...) or .ndjson exports, this release prevents failures caused by the wrong dataset folder structure and ensures the trainer checks the right dataset type.
• More informative classification outputs 🔍
  • Top-5 summaries make it easier to evaluate “near-miss” predictions and build UI/analytics around multiple candidates (watch for potential duplicated entries in Results.summary() output in this specific tag) ⚠️
• More predictable experiment output paths 🧰
  • Workflows that need custom output directories (sweeps, notebooks, CI pipelines) can now set save_dir without it being ignored or flagged.
• Better ARM64 export coverage 🧪
  • Potentially catches ARM-specific NCNN export regressions earlier, though it may also surface new CI failures if upstream NCNN issues persist.

✅ If you’re using Ultralytics Platform + classification, v8.4.2 is a “must update”.

What's Changed

• Return Classify top5 in Results summary by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/23215
• Re-enable NCNN ARM64 exports in Tests CIs by @lakshanthad in https://github.com/ultralytics/ultralytics/pull/23214
• Exclude save_dir from argument validation by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/23191
• ultralytics 8.4.2 Fix Platform Classify training by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/23217

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.1...v8.4.2