New Release: Ultralytics v8.4.125

Discussion
, ,
1

:rocket: Ultralytics v8.4.125: Up to 46% Faster YOLO26 Startup

Ultralytics v8.4.125 significantly accelerates initial YOLO26 model loading in fresh Python processes while reducing unnecessary dependency overhead and improving documentation consistency. Existing workflows should continue without major migration changes.

:high_voltage: New Performance Improvements

Initial model loading is now up to 46% faster, thanks to PR #25883 by @pderrenger:

  • TorchVision transforms are imported only when a checkpoint contains them.
  • Matplotlib loads only when semantic-training label plots are requested.
  • Restricted checkpoint loading is substantially faster while retaining compatibility with older PyTorch versions.
  • The common from ultralytics import YOLO import improved from 558 ms to 471 ms.

Testing with official YOLO26 checkpoints on an Apple M5 Pro showed:

Checkpoint Before After
YOLO26n 451 ms 94 ms
YOLO26x 517 ms 162 ms

That translates to approximately 44–46% faster overall model startup in fresh Python processes. :bullseye:

:test_tube: Expanded Checkpoint Validation

Restricted loading was validated across YOLO26 checkpoints for:

  • Detection
  • Classification
  • Segmentation
  • Pose
  • Oriented bounding boxes
  • Depth
  • Semantic tasks

The update also confirms that multiple model types can load sequentially within one process and improves compatibility with internal torch.nn checkpoint classes.

:link: Documentation and Link Improvements

PR #25879 by @pderrenger standardizes canonical Ultralytics URLs across 136 files, covering documentation, examples, source code, configuration, and workflows.

Several outdated or redirected external references were also refreshed, including links related to datasets, CUDA/cuDNN, MediaPipe, and PASCAL VOC. This means fewer redirects and more reliable navigation throughout the project. :compass:

:package: Upgrade to v8.4.125

Install or upgrade with:

pip install -U ultralytics

Then load a YOLO26 model as usual:

from ultralytics import YOLO

model = YOLO("yolo26n.pt")

No major model architecture or training behavior changes are included, so current YOLO26 workflows should continue to operate normally.

:raising_hands: Try It and Share Your Feedback

Explore the complete details in the Ultralytics v8.4.125 release, or review every change in the v8.4.124 to v8.4.125 full changelog.

Give the faster startup a try in your scripts, services, notebooks, and command-line workflows, and let the Ultralytics team and YOLO community know how it performs! :rocket: