New Release: Ultralytics v8.4.179

Discussion
, ,
1

:rocket: Ultralytics v8.4.179: More Reliable Exports and Inference

Ultralytics YOLO v8.4.179 is here, with fixes for RKNN export compatibility, SAM segmentation, and inference correctness. This release also improves validation, benchmarking, and training stability—no new model architecture is introduced.

:package: Export Reliability and Training Stability

  • RKNN compatibility on Rockchip: Caps rknn-toolkit2 below version 2.5 to avoid build failures affecting YOLOv8 and YOLO11 across Rockchip targets, plus YOLO26 on RK3588. The smoke test now uses imgsz=64 to better catch these failures. PR #26727 was contributed by @glenn-jocher.
  • Clearer TensorRT diagnostics: Unsupported INT8 exports with nms=True on JetPack 6 with TensorRT 10.3 now produce an actionable error. PR #26711 was contributed by @JohnnyWilson16.
  • More stable YOLO26 pose training: Prevents sigma underflow from disabling the pose RLE loss. PR #26713 was contributed by @Nikhi00718.
  • Graceful compilation fallback: Skips torch.compile on CUDA systems when Triton is unavailable. PR #26701 was contributed by @sayaksatpathi.

:bullseye: Segmentation, Validation, and Benchmarking

  • Better SAM segment-everything prompts: Samples points within the image rather than letterbox padding, reducing empty masks, especially on narrow images. Mask resizing is also centralized in the prompt encoder. PR #26708 was contributed by @raimbekovm.
  • Correct RT-DETR class filtering: Removes predictions outside the requested classes before applying the detection limit. PR #26702 was contributed by @Nikhi00718.
  • Fairer open-vocabulary benchmarks: Exported YOLOE and YOLO-World models now use the dataset’s class vocabulary for more consistent comparisons with PyTorch. PR #26704 was contributed by @raimbekovm.
  • Cleaner CPU timing: TorchScript warmup now runs before timed CPU inference. PR #26700 was contributed by @raimbekovm.

:hammer_and_wrench: Correctness and Robustness Fixes

  • Regression safeguards: Fixes cohort regressions and removes an unsafe TorchScript graph rewrite to preserve correct outputs. PR #26698 was contributed by @glenn-jocher.
  • Safer ZIP extraction: Archives containing a root ./ entry no longer trigger an IndexError. PR #26712 was contributed by @Nikhila-Dhevi-R.
  • More resilient mirroring: Failed GitLab mirrors receive one retry before failing. PR #26714 was contributed by @glenn-jocher.

:books: Documentation Updates

  • Ultralytics Platform guidance: Clarifies selectable Roboflow import previews, chart zoom reset, and Platform 409 rules. PR #26706 was contributed by @raimbekovm.
  • ONNX INT8 clarification: Explains what ONNX INT8 export quantizes. PR #26705 was contributed by @AtomicAyush.
  • Additional video resource: Adds a video to the documentation. PR #26709 was contributed by @RizwanMunawar.

:raising_hands: Thanks to Our Contributors

Thanks to the YOLO community and Ultralytics team for these improvements! A special welcome to first-time contributors @JohnnyWilson16, @AtomicAyush, @Nikhila-Dhevi-R, and @sayaksatpathi.

:test_tube: Try v8.4.179

Install this release with:

pip install --upgrade ultralytics==8.4.179

Explore the v8.4.179 release notes for the release details, or review the full changelog to inspect the changes.

Give it a try and share your feedback below! If you encounter an issue, include your environment, model, and a minimal reproduction to help the community investigate.