New Release: Ultralytics v8.4.20

Discussion
, ,
1

Ultralytics v8.4.20 is out :tada::rocket:

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 Ultralytics YOLO models and the Ultralytics Platform.

:glowing_star: Highlights (what you’ll notice first)

:broom: Ray Tune tuning cleanup (most important)

  • Removed hardcoded tuner_callbacks and decoupled built-in W&B callback wiring from RunConfig, making tuning runs less fragile—especially in mixed/distributed environments.
  • See PR #23772 by @Laughing-q

:package: RKNN export reliability improved (Rockchip)

  • Added an ONNX version guard (onnx<1.19.0) and enforced an ONNX opset cap (<=19) for RKNN conversion.
  • Smoother export paths for YOLO26 models on Rockchip toolchains.
  • See PR #23802 by @lakshanthad

:bullseye: FastSAM prompt accuracy fix

  • CLIP prompting now masks non-target neighboring segments before scoring, reducing false positives in overlapping/contained regions.
  • See PR #23766 by @orcinus (first contribution!)

:sparkles: Improvements

:hammer_and_wrench: ByteTracker consistency update

  • Optional score fusion now also applies in second-stage association when fuse_score=True, aligning behavior across tracking stages.
  • See PR #23771 by @ahmet-f-gumustas

:white_check_mark: Better YAML error messages

  • YAML load failures now return clearer syntax errors with file context and validation guidance (instead of ambiguous fallback failures).
  • See PR #23767 by @glenn-jocher

:robot: Jetson / JetPack 6 stack refresh

  • Updated Jetson Docker/runtime stack to newer Torch/Torchvision/ONNX Runtime GPU versions.
  • Compatibility checks now recognize torch 2.10 + torchvision 0.25.
  • See PR #23788 by @lakshanthad and PR #23801 by @lakshanthad

:books: Docs + Ultralytics Platform updates

If you’re new to it, the Ultralytics Platform is the simplest way to annotate, train (cloud or local), deploy (all export formats), and monitor models—see the Ultralytics Platform documentation as you explore.

:bug: Maintenance / CI

:up_arrow: How to get v8.4.20

pip install -U ultralytics

Quick sanity check:

from ultralytics import YOLO

model = YOLO("yolo26n.pt")
model.predict("https://ultralytics.com/images/bus.jpg")

:link: Release + changelog

You can review the full release details in the Ultralytics v8.4.20 GitHub Release and browse changes in the v8.4.19 → v8.4.20 full changelog comparison.

:speech_balloon: Try it and tell us what you think!

If you hit anything odd (tuning/export/Jetson configs in particular), please share:

  • your OS + Python version
  • ultralytics version
  • exact command or minimal repro snippet

Feedback helps keep YOLO sharp for everyone in the community :raising_hands: