Ultralytics v8.4.149 Released
Ultralytics v8.4.149 improves training reliability, export compatibility, and Platform documentation, with important fixes for resumed runs, third-party integrations, and RT-DETR LiteRT accuracy.
Review the complete update in the v8.4.149 release notes.
New Features
SAM 3.1 is now the default smart annotation model
Ultralytics Platform now lists SAM 3.1 as its default smart annotation model, while SAM 3 and SAM 2.1 remain available.
Documentation also covers SAM 3.1 image prediction and current limitations, including unsupported Object Multiplex video tracking.
This update was contributed by @glenn-jocher in PR #26145.
AutoTrain documentation reorganized
AutoTrain documentation has moved to Platform → Train → AutoTrain, with clearer guidance covering:
- Baseline creation
- Controlled comparisons
- Experiment review
- Billing
- Team workspace usage
The documentation update was contributed by @JaviChulvi in PR #26146.
Training and Integration Improvements
More reliable checkpoint resume behavior
Resumed training now preserves the dataset stored in the checkpoint instead of unexpectedly switching to the default COCO8 dataset.
Additional improvements include:
- An explicitly supplied
data=argument still takes priority. - Missing datasets produce the expected dataset-not-found error.
- Checkpoints created with AMP disabled no longer attempt to load an empty AMP scaler state.
Safer ONNX CUDA fallback
CUDA availability is now checked after ONNX Runtime creates its inference session. When the CUDA execution provider is unavailable, inference falls back to CPU with a clear warning instead of crashing.
Improved Ray Tune compatibility
The Ray Tune integration now:
- Verifies that a compatible Ray version is installed.
- Reports training metrics through the current Ray Tune API.
- Reduces integration failures with newer Ray releases.
More robust Weights & Biases runs
Automatically generated W&B run names are now limited to 64 characters, preventing crashes caused by overly long names.
Clearer invalid-argument errors
CLI help is logged separately from syntax errors, helping the core error remain visible when output is truncated.
These reliability improvements were contributed by @glenn-jocher in PR #26148.
Bug Fixes
RT-DETR LiteRT export and inference accuracy
RT-DETR outputs are no longer incorrectly coordinate-normalized during LiteRT export or inference.
This brings LiteRT detection metrics much closer to ONNX results and improves consistency across Python, JavaScript, and npm deployment workflows.
The fix was contributed by @onuralpszr in PR #26142.
Upgrade Today
Install or upgrade to the latest release with:
pip install -U ultralytics
Verify the installed version with:
yolo checks
For every commit included in this release, explore the full v8.4.148 to v8.4.149 changelog.
Give Ultralytics v8.4.149 a try and share your results, questions, or feedback with the YOLO community. Thanks to all contributors and users helping make each release more reliable! 