Ultralytics v8.4.133: Faster Inference, Smarter Tuning, and Richer Metrics
Ultralytics v8.4.133 is now available! This release delivers better hyperparameter tuning convergence, faster inference preprocessing, automatic CPU optimizations, more reliable INT8 calibration, and size-specific detection metrics. Edge-device installation and Weights & Biases artifact management are also simpler. 
Review the official Ultralytics v8.4.133 release for the complete release details.
New Features
Smarter hyperparameter tuning
PR #25984 from @glenn-jocher significantly improves tuning convergence:
- Selects complete, high-performing configurations using fitness weighting.
- Preserves useful relationships between hyperparameters.
- Mutates parameters in normalized search-space coordinates, allowing zero-valued parameters such as
degreesandshearto evolve effectively. - Gradually reduces mutation size when improvements slow down.
- Prevents duplicate candidates in small or discrete search spaces.
- Uses Optuna multivariate TPE with parallel-aware suggestions as the Ray Tune default.
These changes help searches explore broadly before refining the most promising configurations.
Size-specific mAP for custom datasets
PR #25981 from @fcakyon adds small-, medium-, and large-object mAP reporting for custom detection datasets when save_json=True.
Temporary COCO-format annotations are generated internally while preserving native metrics and prediction files. The feature works during training validation, final-model validation, and standalone validation. 
Performance Improvements
Faster predictor preprocessing
PR #25982 from @jahsef moves image channel reordering and tensor-contiguity operations from CPU-side NumPy processing to the inference device.
This preserves output values while reducing unnecessary CPU copies. Reported RTX 5080 benchmarks show approximately 2.2–3.1× faster preprocessing, with additional gains observed on CPU.
Automatic channels-last CPU inference
PR #25983 from @JESUSROYETH automatically enables channels-last memory layout for native PyTorch inference and standalone validation on supported x86 Linux and Windows CPUs with oneDNN.
Training defaults and unsupported platforms remain unchanged, while explicit channels_last=True continues to support compatible CPU and CUDA workflows. Saved models are restored to a safe contiguous format, and stale EMA data is cleared for improved compatibility.
Bug Fixes and Reliability
Correct INT8 calibration fractions
PR #25978 from @JESUSROYETH fixes fraction handling for classification and detection INT8 export calibration.
Scalar fractions now apply directly to the selected calibration split, while list-based fractions retain train, validation, and test behavior. This prevents unexpectedly calibrating against an entire dataset when only a subset was requested.
Improved W&B artifact control
Initial W&B artifact controls arrived through PR #25979 from @fcakyon, followed by integration with the standard training save argument in PR #25985 from @glenn-jocher.
Using save=False now skips uploading the best checkpoint while retaining metrics and plots. The default behavior remains unchanged with save=True.
Simpler Edge-Device Setup
PR #25977 from @Y-T-G updates Raspberry Pi, Jetson, DGX Spark, DeepStream, and related installation guides to use the base ultralytics package instead of the larger [export] extra.
Export dependencies are installed automatically when needed, helping reduce installation size and avoid dependency conflicts on edge devices.
Upgrade Today
Install or upgrade to v8.4.133 with:
pip install -U ultralytics
You can inspect every update in the full v8.4.132 to v8.4.133 changelog.
Give the new release a try and share your benchmarks, feedback, and any issues you encounter. Thanks to every contributor and the wider YOLO community for continuing to improve Ultralytics YOLO! 