New Release: Ultralytics v8.4.66

Discussion
, ,
1

Ultralytics v8.4.66 is out :tada:

A new Ultralytics YOLO release has landed: v8.4.66. This update focuses on practical quality improvements across three important areas: more reliable NVIDIA GPU monitoring :battery:, better semantic segmentation filtering and metrics :brain:, and sharper CoreML semantic exports on iPhone and Apple devices :mobile_phone:.

If you’re building with Ultralytics Platform or working locally with YOLO26, this release should make monitoring, evaluation, and mobile deployment smoother without changing your core workflow.

Quick summary

Highlights in v8.4.66

  • :rocket: More reliable built-in NVIDIA GPU monitoring
  • :bullseye: Proper classes filtering for semantic segmentation
  • :chart_increasing: More accurate semantic segmentation mIoU reporting
  • :mobile_phone: Full-resolution CoreML semantic maps for better iPhone results
  • :gear: Cleaner TensorRT 11 export flow
  • :memo: Clearer docs across YOLO26 testing and training guides
  • :hammer_and_wrench: More robust CI and packaging setup

New features

:battery: More reliable NVIDIA GPU monitoring

GPU stats collection is now more dependable thanks to direct use of NVIDIA’s NVML library.

  • nvidia-ml-py is now included as a standard dependency
  • Built-in GPU reporting no longer relies on runtime install attempts
  • Utilities like auto device selection and system logging should behave more consistently on NVIDIA systems

This improvement comes from PR #23922 by @glenn-jocher from @glenn-jocher.

:bullseye: Better semantic segmentation class filtering

Semantic segmentation now handles the classes filter much more intuitively across the workflow.

This includes:

  • dataset loading
  • prediction
  • validation
  • metric reporting

Non-selected classes are now properly ignored instead of leaking into results, and unsupported configurations like single_cls=True for semantic segmentation are handled more safely.

This feature was added in PR #24806 by @lmycross from @lmycross.

:mobile_phone: Full-resolution CoreML semantic exports

CoreML exports for semantic segmentation now preserve full-resolution semantic maps rather than lower-resolution stride-8 maps.

That means:

  • sharper overlays
  • less blocky masks
  • better visual quality on iPhone and other Apple devices

This landed in PR #24799 by @glenn-jocher from @glenn-jocher.

Improvements

:chart_increasing: More accurate semantic segmentation metrics

Semantic segmentation evaluation is now more faithful to the actual data being evaluated.

  • mIoU now ignores classes not present in the ground truth
  • per-class summaries better reflect what was truly measured

This makes reported results more trustworthy, especially when working with filtered or sparse class sets.

:gear: Cleaner TensorRT 11 export path

TensorRT 11 export was simplified for cleaner precision handling.

  • removed the separate half argument from ONNX precision conversion
  • simplified FP16 and INT8 export flow
  • streamlined INT8 calibration preprocessing

This improvement comes from PR #24795 by @Laughing-q from @Laughing-q.

:bar_chart: Refreshed iPhone CoreML benchmarks

The iPhone CoreML benchmark docs were updated using optimized profile-mode measurements, giving a better picture of real YOLO26 CoreML behavior, especially for semantic models.

See PR #24804 by @glenn-jocher from @glenn-jocher.

Documentation updates

This release also includes several documentation improvements to make testing, training, and integrations easier to follow.

Bug fixes and maintenance

:hammer_and_wrench: CI reliability improvements

To reduce flaky CI failures, dependency installs now automatically retry.

That change is included in PR #24809 by @glenn-jocher from @glenn-jocher.

:package: Packaging cleanup

Because nvidia-ml-py is now a standard dependency, extra manual installs were removed from CI and Docker-related setup, helping keep environments cleaner and more predictable.

Why this release matters

v8.4.66 is a quality-focused release with real day-to-day impact:

  • :white_check_mark: fewer GPU monitoring issues on NVIDIA systems
  • :artist_palette: better usability for semantic segmentation workflows
  • :straight_ruler: more trustworthy segmentation metrics
  • :red_apple: sharper mobile semantic outputs with CoreML
  • :high_voltage: simpler TensorRT 11 export behavior
  • :books: clearer guidance for testing and training with YOLO26

If you’re starting a new project, YOLO26 remains the recommended choice. YOLO11 is still fully supported, but YOLO26 is smaller, faster, more accurate, and natively end-to-end.

Full changelog

You can review the complete diff in the full changelog for v8.4.66, and the packaged release is available on the v8.4.66 release page.

Try it out and let us know what you think :speech_balloon:

We’d love your feedback on this release, especially if you’re using:

  • semantic segmentation
  • CoreML exports
  • TensorRT 11 deployment
  • NVIDIA-based training or benchmarking

Thanks to all contributors who helped ship Ultralytics v8.4.66 :raising_hands: