New Release: Ultralytics v8.4.21

Discussion
, ,
1

:rocket: Ultralytics v8.4.21 is out!

Ultralytics v8.4.21 is a reliability-focused release with a big win for Rockchip RKNN exports :white_check_mark:, plus improved Ray Tune trial isolation, expanded ONNXRuntime C++ pose example support, and clearer Ultralytics YOLO (YOLO26) optimizer guidance. :bullseye:

If you deploy on edge, tune lots of runs, or train YOLO26 regularly, this one’s worth grabbing.

:star: New Features & Key Changes

:white_check_mark: Rockchip RKNN export path fix (top priority)

PR #23806 by @Laughing-q

  • RKNN filename generation now uses safer path handling (Path(...).stem) instead of fragile string replacement.
  • Output naming is cleaner and more consistent, especially with unusual filenames/paths.
  • Version bump 8.4.20 → 8.4.21 included.

Impact: More dependable RKNN export workflows for Rockchip deployment :package:

:test_tube: 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 across trials → improved reproducibility.

Impact: More stable tuning loops and fewer “mystery” inconsistencies :repeat_button:

:person_standing: ONNXRuntime C++ example: pose support (YOLOv8-style pose)

PR #23786 by @chendao12138

  • Adds pose tensor parsing, keypoint decode/scale, NMS for pose results, and visualization.
  • Supports FP32 and FP16 in the example pipeline.
  • Note: YOLO26 isn’t supported yet in this specific C++ example flow.

Impact: Easier starting point for C++ pose inference projects :gear:

:blue_book: Documentation Improvements (YOLO26 training)

:brain: Clearer MuSGD guidance with optimizer=auto

Improves clarity on when MuSGD is selected and when to try it vs standard optimizer options.

Impact: Faster, more confident training decisions for YOLO26 users :control_knobs:

:puzzle_piece: What’s Changed (PRs)

:new_button: New Contributors

Big thanks to first-time contributors (welcome!) :raising_hands:

:up_arrow: Upgrade & Try It

Update with:

pip install -U ultralytics

Then kick the tires on your workflow (especially RKNN export and Ray Tune). If you hit anything odd, please share details in Discussions—your feedback helps keep Ultralytics YOLO solid for everyone. :folded_hands:

You can review the full release details in the v8.4.21 GitHub release and browse the full changelog comparison.