New Release: Ultralytics v8.4.27

Discussion
, ,
1

Ultralytics v8.4.27 is out :rocket:

Quick summary: Ultralytics v8.4.27 brings more reliable Ultralytics Platform training control, especially around cancellation handling :white_check_mark:, along with important fixes for COCO conversion, mask/coordinate alignment, Paddle compatibility, and Docker runtime updates.

If you’re training, exporting, or deploying with Ultralytics YOLO, this release should make workflows a bit smoother and more dependable.

Highlights

:red_circle: Improved Ultralytics Platform training integration

The biggest update in v8.4.27 is a reliability pass on Platform training control in PR #24008 by @glenn-jocher.

This includes:

  • A new sanitizer for callback payloads before sending them to the Ultralytics Platform
  • Safe conversion of non-JSON float values to prevent callback encoding failures
  • Centralized cancellation handling through shared helper logic
  • Consistent cancellation behavior across both training startup and epoch-end callbacks
  • Updated callback reference docs for the new helper flow

Why it matters: cloud-managed training runs are now more predictable, with fewer tracking failures and cleaner cancellation behavior :stop_button:

Improvements

:puzzle_piece: More robust COCO conversion

PR #23998 by @raimbekovm fixes convert_coco() crashes when annotations are incomplete or missing fields like keypoints or segmentation.

Impact: fewer hard crashes when converting imperfect datasets, and better synchronization of annotation lists to avoid IndexError.

:bullseye: Better coordinate and mask alignment

PR #23995 by @raimbekovm updates pad rounding in scale_coords() and scale_masks() to better match LetterBox.

Impact: reduced 1-pixel misalignment issues, especially helpful for segmentation and pose outputs.

:panda: Safer Paddle compatibility

PR #23997 by @Laughing-q tightens the Paddle dependency to paddlepaddle<3.3.0 for export and inference paths.

Impact: better protection against known breakages in newer Paddle versions.

:spouting_whale: Docker runtime refresh

PR #23991 by @glenn-jocher updates the Docker base image to PyTorch 2.11.0 with CUDA 12.8 and cuDNN 9.

Impact: a more modern container baseline with minimal GPU stack disruption.

Docs

:books: Better YOLO26 guide discoverability

PR #23980 by @raimbekovm adds improved cross-links to the YOLO26 End-to-End Detection guide from FAQ and argument docs.

If you’re starting a new project, YOLO26 remains the recommended model family for all use cases, with support across detect, segment, classify, pose, and OBB.

Full list of changes

Try it out :sparkles:

Upgrade with:

pip install -U ultralytics

You can explore the release on the v8.4.27 release page or review the full changelog.

As always, please give it a try and let us know how it works for your training, export, and deployment workflows. Feedback from the community helps the Ultralytics team keep improving YOLO for everyone :raising_hands: