New Release: Ultralytics v8.3.194

[glenn-jocher]

Ultralytics v8.3.194 — smoother exports, CoreML NMS, and non‑blocking telemetry

Quick summary: This release focuses on export stability and visibility, CoreML NMS inference in Python, and a telemetry refactor for clean, background event handling—plus a handful of quality‑of‑life fixes and docs updates. As always, Ultralytics YOLO defaults to YOLO11 for the best all‑around experience.

Summary

Ultralytics v8.3.194 delivers:

• More reliable ONNX and TensorFlow exports with improved simplification and clearer logs.
• CoreML models with embedded NMS that run directly in Python.
• A refactored, non‑blocking Events system that’s independent of Ultralytics HUB.
• Robust segmentation for dense scenes and aligned SAM‑2 examples in docs.
• Faster CI feedback for contributors.

You can review the highlights below and explore the details in the v8.3.194 release notes.

New Features

• CoreML NMS support in Python
  • Export models with NMS embedded and run inference directly—no manual post‑processing for supported detection use cases. See the PR Support CoreML inference with NMS embedded models by @Y-T-G.
  • Example:

    yolo export model=yolo11n.pt format=coreml nms=True
    yolo predict model=yolo11n.mlpackage
    

Improvements

• Exporter dependency and logging
  • Upgraded ONNX simplifier to onnxslim>=0.1.67 for better ONNX and TensorFlow SavedModel exports, restored urllib3 logs for clearer troubleshooting, and kept sentry_sdk quiet. See Bump onnxslim>=0.1.67 in Exporter by @onuralpszr.
• TensorFlow export stability
  • Pinned versions for export to align with onnx2tf: tensorflow>=2.0.0,<=2.19.0 and tf_keras<=2.19.0. Details in Pin tf_keras<=2.19.0 by @Laughing-q.
• Telemetry refactor
  • Moved anonymous Events to ultralytics.utils.events, running in a background thread and independent of HUB. See Refactor Events class by @glenn-jocher.
  • If you import Events directly, update your code:

    from ultralytics.utils.events import events, Events
    

• Segmentation robustness
  • Prevented crashes from OpenCV’s 512‑channel cv2.resize limit by splitting large mask tensors. See Handle cv2.resize 512-channel limit by @ShuaiLYU.
• SAM‑2 API/doc consistency
  • Standardized to source and renamed img to im in inference, with updated examples. See Fix SAM2DynamicInteractivePredictor example by @Y-T-G.

Developer Experience

• Faster CI feedback
  • Reduced GPU CI timeout from 6 hours to 20 minutes to prevent runaway jobs. Details in Prevent long-running billed GPU CI by @glenn-jocher.

Docs Refresh

• New Events reference and updated queue guide
  • Explore the new Events utils reference and the refreshed Queue Management guide featuring a broader‑use tutorial video.
• Content updates
  • The documentation now includes a useful training queue walkthrough video, added in the PR Add YouTube video to docs by @RizwanMunawar.

Why it matters

• More reliable exports and fewer dependency conflicts thanks to the updated onnxslim and clearer logging.
• Smoother Apple workflows with CoreML NMS models that run directly in Python.
• Predictable TensorFlow export behavior via version pinning.
• Cleaner, non‑blocking telemetry that won’t slow your training or export pipelines.
• Fewer runtime errors in dense segmentation scenarios.

How to upgrade and try it

• Upgrade:
  • pip install -U ultralytics
• Quick test with recommended defaults:
  • yolo predict model=yolo11n.pt source='path/to/images'

YOLO11 is the latest stable and recommended model for most scenarios, offering strong accuracy‑speed trade‑offs across detection, segmentation, pose, and classification.

PR Credits

• Pin tf_keras<=2.19.0 by @Laughing-q
• Support CoreML inference with NMS embedded models by @Y-T-G
• Split large channel masks to handle cv2.resize 512 limitations by @ShuaiLYU
• Fix SAM2DynamicInteractivePredictor example in docs by @Y-T-G
• Prevent long-running billed GPU CI by @glenn-jocher
• Refactor Events class by @glenn-jocher
• Add training queue video to docs by @RizwanMunawar
• Bump onnxslim>=0.1.67 in Exporter by @onuralpszr

Explore everything that changed in the full changelog diff between v8.3.193 and v8.3.194.

We’d love your feedback

Please upgrade, try the new CoreML NMS workflow, and let us know how exports and telemetry behave in your environment. Share your thoughts and issues in Discussions, and feel free to open PRs with improvements. Your feedback helps the YOLO community and the Ultralytics team keep pushing forward.