v8.4.127 - Load exported YOLO models with the correct task across all 20 formats (#25886)

🌟 Summary

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

📊 Key Changes

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

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

• Improved OpenVINO inference reliability

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

• Safer training resume behavior

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

• More robust result handling

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

• Expanded CoreML export support

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

• Tracking and YOLOE fixes

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

• Checkpoint and training reproducibility improvements

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

• More direct Ultralytics Platform dataset access

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

• Documentation media delivery improvements

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

🎯 Purpose & Impact

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

What's Changed

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

New Contributors

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

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

v8.4.126 - Make restricted checkpoint loading thread-safe and 40% faster, simplify RLE prior (#25885)

🌟 Summary

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

📊 Key Changes

• 🔒 Thread-safe restricted checkpoint loading

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

• ⚡ Faster restricted model loading

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

• 🛡️ Improved compatibility with secure loading

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

• 🧠 Simplified RLE prior calculation

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

• 🏷️ Version update

  • Updated the Ultralytics package version from 8.4.125 to 8.4.126.

🎯 Purpose & Impact

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

What's Changed

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

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

v8.4.125 - Speed up initial model loading (#25883)

🌟 Summary

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

📊 Key Changes

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

🎯 Purpose & Impact

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

What's Changed

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

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

v8.4.124 - Restore dynamic image sizes for NMS exports (#25874)

🌟 Summary

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

📊 Key Changes

• Dynamic NMS exports restored — PR #25874

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

• Improved CoreML attention export

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

• More efficient RT-DETR training and TensorRT inference

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

• Training reliability fixes

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

• Prediction and result-processing improvements

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

• Export and platform updates

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

🎯 Purpose & Impact

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

What's Changed

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

New Contributors

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

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

v8.4.123 - Accept existing depth dataset formats (#25859)

🌟 Summary

Ultralytics 8.4.123 expands depth-estimation dataset compatibility, allowing standard scaled PNG and floating-point NPY depth maps to be used directly across training, NDJSON conversion, and Ultralytics Platform workflows. 🚀

📊 Key Changes

• 🗂️ Broader depth format support

  • Accepts plain 16-bit grayscale PNG depth maps.
  • Adds support for floating-point .npy depth maps with values stored in meters.
  • PNG files no longer require Ultralytics-specific embedded metadata.

• ⚙️ Configurable depth scaling

  • Adds the optional depth_scale dataset YAML field.
  • Defaults to 1000, meaning PNG values are interpreted as millimeters.
  • Supports datasets with other conventions, such as KITTI (256) and Virtual KITTI 2 (100).

• 🔗 Improved dataset pairing and validation

  • Depth files are matched by filename stem in parallel images/ and depth/ directories.
  • PNG files are preferred, with automatic fallback to NPY files.
  • Depth maps may use a smaller resolution than RGB images when their aspect ratios match.
  • Invalid values, including zero, NaN, and infinity, are safely handled.

• 📥 NDJSON depth dataset support

  • Depth records now only require a paired depth.url.
  • Dataset-level depth_scale is preserved when converting NDJSON to YOLO format.
  • Large downloads are processed in batches to reduce memory usage.

• ☁️ Ultralytics Platform integration

  • Depth datasets can now be uploaded, exported, and used for training on the Platform.
  • NDJSON exports include the depth task, scaling configuration, and paired depth URLs.
  • Depth estimation is now fully listed among the Platform’s supported task types.

• 🧰 Dataset configurations updated

  • Built-in depth datasets now preserve their native storage scales instead of converting everything to meter-based metadata PNGs.
  • Documentation and tests were updated for ARKitScenes, DIODE, KITTI, TartanAir, Virtual KITTI 2, Depth8, and other depth datasets.

🎯 Purpose & Impact

• ✅ Easier dataset adoption: Existing depth datasets can be used with less preprocessing and fewer custom conversion scripts.
• 📦 Simpler, more portable files: Plain PNG and NPY formats work with common tools and do not depend on special PNG metadata.
• 🎯 More accurate dataset handling: Dataset-specific scales preserve the intended precision and depth range.
• 🚀 End-to-end depth workflows: Users can now prepare, upload, convert, and train depth datasets through the Ultralytics Platform.
• ⚠️ Migration consideration: Older self-describing Ultralytics depth PNGs that rely on embedded metadata may need to be converted to the new scaled PNG format. See the depth dataset format documentation.

What's Changed

• Accept existing depth dataset formats by @glenn-jocher in #25859

Full Changelog: v8.4.122...v8.4.123

v8.4.122 - Use canonical PNG depth maps (#25858)

🌟 Summary

v8.4.122 standardizes Ultralytics depth estimation around compact, self-describing 16-bit PNG depth maps—simplifying the data pipeline and significantly reducing storage requirements. 🖼️📉

📊 Key Changes

• Canonical depth format introduced: Depth targets now use 16-bit PNG files with embedded metadata describing their meter range.
• Invalid pixels are clearly represented: Pixel value 0 is reserved for missing or invalid depth values, allowing training and evaluation to ignore them safely.
• New depth utilities added:
  • save_depth_png() writes depth maps in the required format.
  • load_depth() reconstructs meter-valued depth maps for training and inference.
• NPY support removed: The depth loader, converters, manifests, path handling, and compatibility fallbacks no longer support .npy depth maps.
• Dataset support updated: Built-in depth dataset configurations and generators—including Depth8, NYU Depth, KITTI, SUN RGB-D, DIODE, Hypersim, TartanAir, Virtual KITTI 2, and ARKitScenes—now produce and consume PNG depth maps.
• NDJSON conversion updated: Depth records must use the linear-u16 PNG encoding contract.
• Documentation and tests refreshed: Depth dataset guides, task documentation, reference pages, and validation tests now describe the PNG-only workflow.
• Smaller published archives:
  • Depth8 archive reduced to approximately 8.9% of its previous size.
  • NYU Depth archive reduced to approximately 39.6% of its previous size. 🚀

🎯 Purpose & Impact

• Lower storage and download costs: PNG compression substantially reduces dataset and hosted asset sizes.
• More portable depth files: Each PNG contains the information needed to interpret its depth values, making files easier to move, inspect, and process independently.
• Simpler, more consistent tooling: All supported depth datasets now follow one file format and one loading path.
• Improved browser/display behavior: The encoding preserves valid depth information when 16-bit assets are viewed through systems that reduce them to 8-bit displays.
• Important migration requirement: Existing custom datasets using .npy depth maps must be converted to the new metadata-enabled PNG format before use with this release. Existing hosted depth assets also require the planned production migration; no runtime compatibility layer is included.
• No major model architecture change: This release primarily improves the YOLO26-Depth data representation and dataset pipeline, rather than changing model structure or training objectives.

What's Changed

• Use canonical PNG depth maps by @glenn-jocher in #25858

Full Changelog: v8.4.121...v8.4.122

v8.4.121 - Fix OpenVINO INT8 detection head scope (#25841)

🌟 Summary

v8.4.121 improves OpenVINO INT8 export reliability for YOLO26 models while delivering broad Platform API, dataset, annotation, deployment, and documentation updates. 🚀

📊 Key Changes

• Fixed OpenVINO INT8 detection-head handling by @glenn-jocher:

  • Replaced fragile PyTorch-based layer matching with exact names from the converted OpenVINO graph.
  • Keeps Detect decoding, DFL, and Sigmoid operations in floating point as intended.
  • Preserves strict NNCF validation during quantization.
  • Verified with a successful YOLO26n-P2 INT8 export using 55 exact ignored operations, with no unwanted FakeQuantize nodes. ✅
  • This directly addresses export failures reported in Sentry and is the most important change in this release.

• Expanded Ultralytics Platform API documentation:

  • Documented the generated ultralytics-platform Python SDK alongside REST examples.
  • Updated endpoint paths, authentication, pagination, rate limits, response formats, and OpenAPI guidance.
  • Added coverage for images, dataset ingestion, exports, storage integrations, billing, usage, trash, training, deployments, and account APIs.
  • Clarified that workspace API keys have owner-level permissions and are managed by workspace owners.

• Improved Platform dataset and annotation workflows 🏷️:

  • Added clearer documentation for URL, cloud-storage, and On Premise dataset imports.
  • Documented class merging/deletion, conflict handling, dataset readiness checks, clustering, version restore, and expanded annotation controls.
  • Added support documentation for COCO and NDJSON imports, while clarifying that Pascal VOC XML labels are not imported.
  • Documented annotation visibility controls, copy/paste workflows, and new keyboard shortcuts.

• Updated Platform account, billing, and team documentation 💳:

  • Added the Usage tab, detailed credit metering, monthly credit expiration, auto top-up behavior, seat billing, renewals, and downgrade effects.
  • Clarified workspace roles, owner-only API keys, team invitations, seat reuse, ownership transfer, and team deletion.
  • Expanded activity exports, trash permissions, storage usage, and account deletion guidance.

• Improved deployment and inference documentation 🌐:

  • Clarified dedicated endpoint lifecycle operations, authentication, rate limits, model replacement, health checks, metrics, logs, and capacity behavior.
  • Documented video inference, endpoint-specific API references, depth response options, and generated deployment URLs.

• Security and CI improvements 🔒:

  • Prevented checkout credentials from being copied into Docker images.
  • Moved workflow secrets into environment variables instead of embedding them in scripts.
  • Updated self-hosted runner cleanup actions to v1.4.39.
  • Reduced individual SlowTests attempts from 180 to 120 minutes while retaining one retry.
  • Changed Dependabot GitHub Actions checks from daily to weekly.

• Dependency and documentation maintenance 📚:

  • Bumped the package version to 8.4.121.
  • Allowed newer setuptools and NNCF versions, including NNCF 3.x.
  • Updated Rust inference examples to ultralytics-inference 0.0.35.
  • Corrected task banners and documentation to consistently represent YOLO26’s supported task coverage.

🎯 Purpose & Impact

• More dependable edge deployment: YOLO26 users exporting INT8 models through OpenVINO should see fewer conversion failures and more predictable quantization behavior. ⚡
• Better model accuracy preservation: Keeping detection-head decode operations in floating point helps protect inference correctness while still applying INT8 compression elsewhere.
• Easier Platform automation: The updated REST contract and generated Python SDK documentation make it simpler to build integrations, manage datasets, start training, create exports, and operate deployments.
• Clearer team and billing behavior: Users can better understand permissions, shared workspace costs, credit expiration, seat charges, and training metering before taking action.
• More capable dataset preparation: New import, annotation, class-management, clustering, and versioning guidance supports more complete end-to-end computer vision workflows.
• Safer builds and more stable CI: Docker images are less likely to contain authentication credentials, while CI jobs are better protected against hangs and stale runner state. 🛡️

What's Changed

• Trainer by @AyushExel in #24
• Prevent checkout credentials from entering Docker images by @glenn-jocher in #25805
• Bump eviden-actions/clean-self-hosted-runner from v1.4.37 to v1.4.38 in /.github/workflows by @UltralyticsAssistant in #25811
• Update Platform docs to match current product and API contract by @glenn-jocher in #25808
• Bump ultralytics-inference version to 0.0.35 in documentation by @onuralpszr in #25818
• Fix spelling in Platform account settings by @UltralyticsAssistant in #25821
• Document the ultralytics-platform SDK across the Platform docs by @glenn-jocher in #25823
• Standardize .github configuration by @glenn-jocher in #25822
• Update the supported tasks banner and the task enumerations by @raimbekovm in #25807
• Bump eviden-actions/clean-self-hosted-runner from v1.4.38 to v1.4.39 in /.github/workflows by @UltralyticsAssistant in #25836
• Update setuptools requirement from <81.0.0 to <85.0.0 by @dependabot[bot] in #25826
• Update nncf requirement from <3.0.0,>=2.14.0 to >=2.14.0,<4.0.0 by @dependabot[bot] in #25825
• Bound SlowTests retry attempts by @glenn-jocher in #25839
• Fix OpenVINO INT8 detection head scope by @glenn-jocher in #25841

Full Changelog: v8.4.120...v8.4.121

v8.4.120 - Avoid nondeterministic CUDA anchor cumsum (#25806)

🌟 Summary

Ultralytics 8.4.120 improves CUDA training determinism and TensorFlow export reliability, while expanding documentation for LLM workflows and AI coding-agent integrations. 🚀

📊 Key Changes

• Deterministic CUDA anchor generation by @glenn-jocher

  • Replaced CUDA cumulative-sum operations with deterministic arange-based generation when creating detection anchors.
  • Removes recurring cumsum_cuda_kernel warnings during deterministic training.
  • Preserves runtime device handling for traced and TorchScript GPU models, avoiding device information being incorrectly fixed during tracing.

• More reliable TensorFlow exports 🛠️

  • Removed the obsolete NVIDIA package index from TensorFlow and non-YOLO export dependency installation.
  • onnx-graphsurgeon can now be installed directly from PyPI, reducing DNS and connectivity issues—especially in CPU-based CI environments and isolated export setups.

• New Ultralytics LLM documentation 🤖

  • Documents the OpenAI-compatible LLM interface for text, image, streaming, asynchronous, provider-specific, and YOLO-combined workflows.
  • Provides examples for OpenAI-compatible services such as DeepSeek, Kimi, Z.AI GLM, OpenRouter, and local servers.
  • Updates the default documented and runtime model to gpt-5.6-luna.

• New Agent Skills integration guide 🧩

  • Documents the official ultralytics/skills repository.
  • Covers AI-agent skills for model selection, datasets, training, tuning, inference, and export.
  • Includes installation guidance for Claude Code, Codex, and other compatible agents.

• Version update

  • Bumped the Ultralytics package version from 8.4.119 to 8.4.120.

🎯 Purpose & Impact

• ✅ Cleaner deterministic training logs: Users no longer see repeated CUDA cumsum warnings that can obscure important training messages.
• ✅ More predictable model tracing: TorchScript and traced GPU models retain runtime device behavior without sacrificing the deterministic anchor-generation fix.
• ✅ Smoother TensorFlow export setup: Fewer external package-index dependencies should improve export reliability in restricted networks, CI pipelines, and CPU-only environments.
• ✅ Better LLM discoverability: Developers can more easily connect YOLO detection results with language and vision models through a consistent interface.
• ✅ Improved AI-assisted development: Agent Skills provide structured, workflow-specific guidance for using Ultralytics tools with supported coding agents.
• ℹ️ No major model architecture changes were introduced in this release; the primary technical improvement is improved determinism and export robustness.

What's Changed

• Document the OpenAI-compatible LLM interface by @onuralpszr in #25789
• Document Ultralytics Agent Skills by @JaviChulvi in #25787
• Avoid nondeterministic CUDA anchor cumsum by @glenn-jocher in #25806

New Contributors

• @JaviChulvi made their first contribution in #25787

Full Changelog: v8.4.119...v8.4.120

v8.4.119 - Enable OpenVINO NPU_TURBO for classification models on supported NPUs (#25784)

🌟 Summary

v8.4.119 improves Intel NPU classification performance with OpenVINO, strengthens detection and tracking reliability, expands Platform integration, and refreshes documentation and developer tooling. 🚀

📊 Key Changes

• ⚡ OpenVINO NPU_TURBO for classification

  • Enables NPU_TURBO automatically for classification models running on supported Intel NPU devices.
  • Applies only when the device advertises support, preserving existing behavior for other tasks, devices, and drivers.
  • Measured classification latency improvements of approximately 42–52% on an Intel Core Ultra 9 185H, with bit-identical accuracy across tested configurations.
  • Turbo mode is intentionally limited to classification because larger image sizes showed little benefit and could consume unnecessary power.

• 🔌 Platform SDK exposed from ultralytics

  • Python 3.11+ users can access Platform, AsyncPlatform, APIError, and APIConnectionError directly from the main package.
  • Imports are lazy, so the SDK is loaded only when needed.
  • Python 3.8–3.10 users receive a clear compatibility message when requesting Platform exports.

• 🛰️ More efficient Platform training callbacks

  • Subsequent metrics, telemetry, and model uploads now reuse the registered model ID instead of repeatedly resolving project and model names.
  • Reduces internal Platform traffic and improves reliability during long training runs.

• 🛡️ Improved tracking robustness

  • BYTETracker now ignores detections with zero or negative width or height before creating tracks, preventing invalid Kalman filter states.
  • Kalman filter operations were simplified to use direct slicing instead of unnecessary matrix operations.
  • Removed redundant array copies in multi-object tracking paths.

• 🎯 Safer detection export postprocessing

  • Detection top-k selection is now limited to the number of available anchors.
  • Prevents export failures on very small input sizes while preserving normal behavior for standard inputs.

• 🧩 Corrected CopyPaste augmentation probability

  • The CopyPaste transform now respects its configured probability in flip mode.
  • Previously, eligible images could receive the augmentation unconditionally.

• 📚 Documentation and usability updates

  • Updated supported-task banners to cover all seven supported tasks, including semantic segmentation and depth estimation.
  • Added the Cmd/Ctrl+Delete image-deletion shortcut to the Platform annotation guide.
  • Added a YOLO26 LiteRT export and mobile deployment tutorial video.
  • Restored light, dark, and system theme controls in the documentation.
  • Removed documentation embeds that did not render correctly on the live site.
  • Improved ASCII string checks using Python’s built-in string operation for faster annotation rendering.

• 🧰 CI maintenance

  • Updated the self-hosted runner cleanup action used across several CI jobs.

🎯 Purpose & Impact

• 🚀 Faster edge inference: Supported Intel NPU users running classification models should see substantially lower latency without sacrificing accuracy.
• 🔋 Better power awareness: Restricting turbo mode to classification avoids enabling a potentially costly optimization where it provides little practical benefit.
• 🤝 Simpler Platform development: Python 3.11+ applications can use Platform SDK functionality through familiar top-level ultralytics imports.
• 📈 More reliable training runs: Model ID reuse reduces repeated lookup work and supports smoother metric streaming and checkpoint uploads.
• 🛡️ Fewer runtime and export failures: Invalid tracking boxes and small-input detection exports are handled safely instead of producing corrupted states or indexing errors.
• 🎛️ More predictable augmentation: CopyPaste now behaves according to its documented probability setting.
• 🌐 Improved user experience: Refreshed documentation, working theme controls, clearer task coverage, and better Platform guidance make the ecosystem easier to use for both new and experienced users.

What's Changed

• Update supported tasks banner by @raimbekovm in #25788
• Document Platform image delete shortcut by @glenn-jocher in #25791
• Use model IDs for Platform training callbacks by @glenn-jocher in #25793
• Bump eviden-actions/clean-self-hosted-runner from v1.4.36 to v1.4.37 in /.github/workflows by @UltralyticsAssistant in #25797
• Add https://youtu.be/FmgrfmZlhpY to docs by @RizwanMunawar in #25795
• Expose Platform SDK from ultralytics by @glenn-jocher in #25800
• Restore dark theme palette in docs configuration by @onuralpszr in #25802
• Replace the per-character is_ascii loop with str.isascii() by @raimbekovm in #25801
• Remove the two docs embeds that do not render on the live site by @raimbekovm in #25792
• fix: add probability guard to CopyPaste in flip mode by @Rahulbiradar9 in #25796
• Clamp detection top-k to available anchors by @Dorablank in #25771
• Replace selection-matrix matmuls with direct slicing in the Kalman filter by @JESUSROYETH in #25770
• Reject zero/negative-height detections before track creation in BYTETracker by @JESUSROYETH in #25769
• Enable OpenVINO NPU_TURBO for classification models on supported NPUs by @synml in #25784

New Contributors

• @Dorablank made their first contribution in #25771

Full Changelog: v8.4.118...v8.4.119

v8.4.118 - Add standalone LLM model interface (#25761)

🌟 Summary

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

📊 Key Changes

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

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

• 📐 Improved oriented bounding box training

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

• ⚡ Faster CopyPaste augmentation

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

• 🧠 More reliable YOLOE behavior

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

• 🏋️ Training and inference stability fixes

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

• 🗂️ Dataset and prediction improvements

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

• 📚 Documentation and deployment updates

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

🎯 Purpose & Impact

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

What's Changed

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

New Contributors

• @doublecurry made their first contribution in #25749

Full Changelog: v8.4.117...v8.4.118