New Release: Ultralytics v8.4.142

Discussion
, ,
1

:rocket: Ultralytics v8.4.142 Is Here!

Ultralytics v8.4.142 delivers a simpler, unified nms API for inference and export, fairer cross-format benchmarks, improved deployment guidance, and a fix for resuming training with Albumentations transforms.

Explore the Ultralytics v8.4.142 release for the complete release details.

:glowing_star: New Features

Unified nms Behavior

The legacy end2end option is now unified under a single nms setting across prediction, validation, tracking, benchmarking, and export:

  • nms=None (default): Uses the one-to-many head with Ultralytics-managed NMS for stronger default accuracy.
  • nms=True: Uses the one-to-many head and embeds NMS in supported export formats.
  • nms=False: Selects the one-to-one, NMS-free head when available for lower-latency deployment and simpler post-processing.

Existing configurations remain compatible:

  • end2end=True maps to nms=False.
  • end2end=False maps to nms=None, unless explicitly combined with nms=True.

Both YOLO26 heads remain trained, while validation, checkpoint selection, and early stopping now follow the selected inference head.

This update was introduced in PR #26066, authored by @glenn-jocher, alongside expanded tests for prediction, validation, export, Hailo, and model-head behavior.

:high_voltage: Improvements

More Reliable Model Fusion and Export

Model fusion now removes the unused detection branch regardless of the selected inference head. Exported models preserve their native output behavior, while unsupported formats automatically fall back to compatible raw outputs.

The nms behavior is now documented consistently across ONNX, TensorRT, CoreML, OpenVINO, MNN, Ascend, Hailo, and other integrations.

:warning: Some runtimes and quantized configurations cannot support NMS-free or embedded-NMS graphs. Check your target integration’s compatibility before deployment.

Fairer Cross-Format Benchmarks

Benchmark validation now uses square inputs consistently for native and exported models. This prevents preprocessing differences from affecting accuracy and latency comparisons.

PR #26068, authored by @glenn-jocher, implements this benchmark consistency improvement.

Expanded Edge Deployment Documentation

This release broadens and refines deployment guidance for specialized hardware:

  • PR #25452, authored by @klemen1999, adds deployment guidance for Luxonis OAK RVC2 and RVC4 cameras.
  • PR #25465, authored by @smohan-ambarella, improves Ambarella CVflow instructions for SDK compilation, host validation, and Cavalry device deployment.
  • PR #26067, authored by @Gornoka, documents reliable DEEPX SDK installation through the vendor wheel repository.

:bug: Bug Fixes

Reliable Training Resume with Albumentations

Resuming training with Albumentations transforms no longer fails while saving arguments to YAML. The fix normalizes augmentation overrides after restoring resume arguments without changing the existing training workflow.

PR #26069, authored by @glenn-jocher, resolves the argument-normalization ordering issue.

:raising_hands: New Contributors

A warm welcome and thank you to our first-time contributors:

:package: Try v8.4.142

Upgrade to the latest release with:

pip install -U ultralytics

For every commit included in this update, review the full v8.4.141 to v8.4.142 changelog and share your feedback, benchmark results, or deployment experiences with the Ultralytics YOLO community. :tada: