New Release: Ultralytics v8.4.82

Discussion
, ,
1

Ultralytics v8.4.82 is out :rocket:

Quick summary: Ultralytics v8.4.82 is a stability-focused release that makes Ultralytics YOLO workflows more reliable, especially for YOLO26 Axelera exports, classification RAM caching, and training/data edge cases. If you’re deploying YOLO26 or training classification models at scale, this update is worth grabbing. :white_check_mark:

You can explore the release on GitHub Releases and review the full changelog.

:glowing_star: Highlights

:rocket: More reliable YOLO26 Axelera exports

This is the biggest fix in v8.4.82.

We fixed an issue where Axelera export could accidentally remove its own generated files when run from the model directory, leading to confusing failures like “output model too small” even when compilation had actually succeeded.

This release:

  • Uses a safer temporary compile directory
  • Prevents cleanup from deleting valid outputs
  • Adds better export serialization to avoid shared-state conflicts
  • Restores support for end2end=False exports where supported

This should make deployment pipelines much more dependable, including automated jobs on the Ultralytics Platform.

:brain: Classification cache='ram' is back, safely

Classification training can now use cache='ram' again without the previous memory growth issue.

Instead of duplicating decoded images across workers, images are now stored in a shared memory buffer, improving memory efficiency while preserving the speed benefits of RAM caching.

:stop_sign: Faster failure for empty-label datasets

Training now stops immediately with a clear error when a detection dataset contains no usable labels at all. This helps catch dataset issues earlier and avoids more confusing failures later in training.

:framed_picture: Better real-world image handling

We improved image format handling by:

  • Adding proper fallback support for .heif
  • Removing the invalid .jpeg2000 extension from supported image lists

:broom: Python reliability improvements

Several subtle mutable default argument bugs were fixed across metrics, TAL, augmentations, YOLOE, and SAM-related code. These changes improve long-running reliability in notebooks, services, and repeated train/infer workflows.

Improvements

:books: Docs and workflow polish

A few helpful quality-of-life updates landed as well:

  • Clearer documentation for stream=True behavior in predict mode
  • Added docs for MLFLOW_KEEP_RUN_ACTIVE for MLflow users
  • Updated citations for YOLOv7, YOLOv9, YOLOv10, and YOLO12 docs
  • Cleaner Slack release and CI notifications

What’s changed

:raising_hands: New contributors

Thanks to our new contributors for jumping in:

Try it out

Update with:

pip install -U ultralytics

If you’re starting a new project, we recommend YOLO26 on Ultralytics Platform for the latest and best-performing Ultralytics model family, with full support for detect, segment, classify, pose, and OBB workflows.

Feedback welcome :speech_balloon:

Please give v8.4.82 a try and let us know how it works in your training and deployment pipelines. Feedback, bug reports, and regression checks are always appreciated from the community.