New Release: Ultralytics v8.4.45

Discussion
, ,
1

:rocket: Ultralytics v8.4.45 Released

Summary: Ultralytics v8.4.45 is a focused hotfix release that restores correct pretrained training behavior for .pt models, improves TensorRT and Jetson export reliability, and adds clearer docs for tuning and the Ultralytics Platform. :tada:

If you train from pretrained checkpoints such as Ultralytics YOLO26, this is an important update.

:fire: Highlights

Critical training hotfix

The biggest fix in v8.4.45 restores expected behavior for training from .pt checkpoints:

  • YOLO("model.pt").train() now correctly starts from pretrained weights again
  • CLI training also respects pretrained checkpoint defaults
  • pretrained=False only disables weight loading when explicitly set
  • Custom pretrained weight paths are preserved
  • Classification scratch-training reset behavior was restored

This addresses a regression that could silently cause training to start without pretrained weights, costing both time and accuracy.

Implemented in PR #24378 by @glenn-jocher.

Better TensorRT and Jetson reliability

This release also hardens deployment workflows:

  • Blocked problematic tensorrt==10.2.0 to avoid known runtime and export issues
  • Unified TensorRT dependency checks with shared check_tensorrt()
  • Improved JetPack 6 INT8 export reliability by upgrading TensorRT from 10.3 to 10.7 in Docker where needed
  • Scoped automatic end2end disabling to the specific problematic combination: JetPack 6 + TensorRT 10.3.0 + INT8
  • Added clearer warnings and troubleshooting guidance

These improvements are especially helpful for Jetson and TensorRT users deploying YOLO models in production.

:new_button: New Features and Improvements

Expanded export argument support

The data argument is now accepted for more export formats, helping dataset-aware INT8 and export workflows across:

  • TensorRT
  • OpenVINO
  • CoreML
  • TFLite
  • TF.js
  • MNN
  • IMX

Added in PR #24362 by @lakshanthad.

Ultralytics Platform docs: Dataset Clustering

New docs were added for Dataset Clustering in the Ultralytics Platform documentation, making it easier to explore clusters, outliers, and duplicates interactively in 2D.

Added in PR #24376 by @sergiuwaxmann.

Hyperparameter tuning docs update

The hyperparameter tuning guide was updated to better match current tuner behavior and outputs, reducing confusion for users tuning custom models.

Updated in PR #24372 by @raimbekovm.

:hammer_and_wrench: Bug Fixes

:bullseye: Why this release matters

This update is small but important:

  • :white_check_mark: Restores expected pretrained training behavior
  • :rocket: Reduces avoidable TensorRT and Jetson deployment failures
  • :toolbox: Improves export workflows with broader data argument support
  • :books: Makes documentation more accurate and useful

For anyone training from .pt checkpoints or exporting models for edge deployment, v8.4.45 is recommended.

:package: Get the release

You can review the full release in the v8.4.45 release page and browse all changes in the full changelog from v8.4.43 to v8.4.45.

:speech_balloon: Try it and let us know

Please give Ultralytics v8.4.45 a try and share your feedback, especially if you use pretrained .pt training, TensorRT export, Jetson, or the Ultralytics Platform.

Thanks to everyone in the YOLO community and to all contributors who helped improve this release! :raising_hands: