New Release: Ultralytics v8.4.33

Discussion
, ,
1

Ultralytics v8.4.33 is out :tada:

Quick summary: Ultralytics v8.4.33 is a stability-focused release with one especially important fix for resumed training in end-to-end YOLO workflows, plus more flexible Ray Tune search support for YOLO26 and a CoreML export improvement for Apple deployment :rocket:

If you train long-running models, tune at scale, or export to CoreML, this update is well worth a look.

Highlights

:wrench: More reliable resume training for end-to-end models

The biggest user-facing improvement in v8.4.33 is a fix to resumed training behavior for end-to-end models.

With PR #24074 by @Laughing-q, resume_training() now correctly restores and synchronizes the model loss criterion when loading from checkpoints. It also restores the internal loss update counters used by one-to-one and one-to-many training, so resumed runs continue with the correct loss progression instead of partially reset behavior.

Why this matters: if a training job is interrupted, resumed training now behaves much more consistently and predictably :white_check_mark:

:brain: More Ray Tune search algorithms for YOLO26

Hyperparameter tuning gets a nice upgrade in PR #23946 by @lmycross.

This update adds search_alg support to model.tune(..., use_ray=True), giving Ultralytics YOLO users more control over tuning strategy. Supported options now include:

  • optuna
  • hyperopt
  • bohb
  • ax
  • nevergrad
  • zoopt
  • random

There are also a few quality-of-life improvements:

  • clearer internal handling around iterations
  • better compatibility for algorithms with specialized search-space requirements
  • automatic use of HyperBandForBOHB when bohb is selected
  • improved docs, tests, and search-space typing

Why this matters: YOLO26 tuning workflows are now more adaptable across different infrastructure setups and optimization preferences :bullseye:

:red_apple: Improved CoreML detection export with NMS

For Apple deployment workflows, PR #24078 by @glenn-jocher improves CoreML detection export when nms=True.

The CoreML pipeline now receives an explicit output shape during export, which helps reduce shape-related issues at export time and runtime on Apple devices.

Why this matters: smoother CoreML deployment with fewer surprises on-device :mobile_phone:

Improvements in this release

  • Fixed progressive loss state restoration when resuming end-to-end training
  • Expanded Ray Tune search algorithm support for YOLO26
  • Improved scheduler behavior for BOHB tuning
  • Refined search-space typing in tuning workflows
  • Increased CoreML export stability for detection models using integrated NMS

Why update?

Ultralytics v8.4.33 is not a flashy feature release, but it delivers meaningful improvements where reliability matters most:

  • Training stability for resumed end-to-end runs
  • Tuning flexibility for advanced Ray Tune users
  • Export robustness for CoreML deployment

In short, this is a strong stability + flexibility update.

Try it out

You can upgrade with:

pip install -U ultralytics

And if you are exploring the latest models, Ultralytics YOLO26 remains the recommended choice for new projects, with support across detect, segment, classify, pose, and OBB workflows on the Ultralytics Platform.

Release links

You can explore the full release in the v8.4.33 release page and review everything included in the full changelog from v8.4.32 to v8.4.33.

What’s changed

Thanks to everyone who contributed to this release :raising_hands: Please give v8.4.33 a try and share feedback, training results, or export experiences with the community.