Ultralytics v8.4.108 Released
Ultralytics YOLO v8.4.108 is now available! This release improves MuSGD stability, inference and tracking efficiency, performance analytics, ONNX Runtime compatibility, Platform integration, and documentation quality—without changing model architecture or accuracy.
Explore the complete v8.4.108 release on GitHub.
Highlights
Fixes MuSGD crashes for models containing LayerScale and other non-matrix parameters.- Reduces tracking overhead while preserving the original detection device.
Adds per-image prediction timing and processed-area reporting.
Improves ONNX Runtime examples for fixed and dynamic input dimensions.
Expands training and inference analytics, including final outcomes and GPU counts.
Connects supported model weights directly to their Ultralytics Platform model pages.
Improves generated API references, examples, dataset links, and documentation reliability.
New Features
Better prediction and training analytics
Predictors now retain per-image preprocessing, inference, and postprocessing times through predictor.speed, along with average processed image area through predictor.pixels. Timing data resets correctly between runs and supports mixed image sizes.
Training events now include final fitness, duration, completed epochs, optimizer, architecture, and dataset details. Multi-GPU runs report the actual GPU count, while invalid task and mode combinations are excluded.
- @glenn-jocher added predictor speed attributes and prediction event fields in PR #25440.
- @glenn-jocher added final training outcomes to analytics in PR #25456.
- @glenn-jocher restricted events to valid modes and tasks in PR #25461.
- @glenn-jocher replaced the distributed-training flag with GPU-count reporting in PR #25463.
Expanded Ultralytics Platform integration
Weights for YOLO26, YOLO11, YOLOv8, and supported YOLOv5 variants now link to their corresponding Platform pages. Huawei Ascend 310B1 and 310B4 targets are also documented as supported.
- @glenn-jocher linked model weights to Platform pages in PR #25446.
- @glenn-jocher updated OrangePi Huawei Ascend 310B Platform support in PR #25459.
Improvements
Faster, more consistent tracking
TrackTrack now calculates motion cues only for detection pairs supported by IoU, reducing unnecessary CPU work in crowded scenes. Redundant NumPy copies were removed, and tracked boxes remain on the same device as their source detections.
- @onuralpszr reduced TrackTrack CPU overhead and preserved source-device placement in PR #25448.
Leaner data augmentation
RandomFlip no longer performs an unnecessary contiguous array copy.
- @JESUSROYETH removed the redundant copy in PR #25462.
More reliable documentation
Generated references now preserve union types such as str | Path, multi-line return descriptions, inherited constructors, example captions, frontmatter, and empty argument tables. Dataset links and automated link checks are also more reliable.
- @glenn-jocher replaced unstable dataset links in PR #25445.
- @glenn-jocher excluded Gitee from automated broken-link checks in PR #25449.
- @glenn-jocher fixed union-type table rendering in PR #25455.
- @glenn-jocher corrected reference truncation, constructor signatures, and empty argument tables in PR #25457.
- @glenn-jocher improved example captions and wrapped table descriptions in PR #25458.
- @glenn-jocher documented inherited constructors and improved frontmatter parsing in PR #25460.
Bug Fixes
MuSGD stability
MuSGD now applies its Muon-style update only to supported 2D linear weights and 4D convolution filters. LayerScale gains, normalization weights, biases, and other parameters continue through the existing SGD update path, preventing AssertionError failures on tensors such as (C, 1, 1).
ONNX Runtime input handling
The YOLOv8 and YOLO11 ONNX Runtime example now handles image height and width correctly during letterbox preprocessing and supports models with dynamic input dimensions through fallback logic.
- @onuralpszr corrected ONNX Runtime input-size handling in PR #25451.
Try It Now
Upgrade to the latest release with:
pip install -U ultralytics
Please test v8.4.108 in your training, prediction, tracking, and export workflows and share your feedback with the community. For every commit included in this release, review the full v8.4.107 to v8.4.108 changelog.