New Release: Ultralytics v8.4.211

Discussion
, ,
1

:rocket: Ultralytics v8.4.211: Faster SAM, INT8 Deployment, and OCR Improvements

Ultralytics YOLO v8.4.211 is here! This release accelerates compiled SAM image encoding, adds static INT8 export for ExecuTorch, and improves OCR, training, CPU inference, and export reliability.

Explore the v8.4.211 release notes for the complete release overview.

:glowing_star: Performance Highlights

  • Faster compiled SAM 3 and SAM 3.1: PR #27158 by @glenn-jocher makes image encoding 7–9% faster in the tested torch.compile setup. Eager-mode performance is unchanged.
  • Faster OCR on NVIDIA GPUs: PR #27143 by @glenn-jocher runs the PP-OCRv6 recognizer with a cached FP16 TensorRT engine, delivering about 2.5× faster recognition than ONNX Runtime on dense pages after the initial engine build.
  • Faster deterministic training: PR #27154 by @raimbekovm accelerates resizing for semantic segmentation and depth workloads, with speedups reaching 3.9× in one semantic segmentation setup.
  • Lower CPU pooling latency: PR #26082 by @glenn-jocher introduces equivalent separable max pooling in SPPF, reducing measured block latency by 22–45% without changing GPU or exported-model behavior.

Actual gains depend on your hardware, model, and settings.

:package: Deployment and Training Improvements

  • Static INT8 export for ExecuTorch: PR #27150 by @raimbekovm adds calibration-based quantization for smaller deployments. Detection, segmentation, pose, and oriented-box heads stay in FP32 to help preserve prediction quality.
  • Classification class weighting: PR #27148 by @raimbekovm extends cls_pw support to classification training, helping account for class imbalance.
  • More efficient CI caching: PR #27156 by @fcakyon enables CPU CI runners to save and restore zstd-compressed GitHub Actions caches.

:hammer_and_wrench: Bug Fixes

  • TensorRT INT8 reliability: PR #26220 by @Y-T-G fixes broken RT-DETR TensorRT 10 INT8 engines and keeps sensitive head output layers in floating point.
  • Pose and classification inputs: PR #27146 by @sohailshk fixes Tiger-Pose keypoint flips and classification prediction on float NumPy arrays.
  • Legacy CoreML exports: PR #27131 by @amanharshx fixes INT8 weight quantization for legacy CoreML exports.

:books: Platform and Documentation Updates

  • Dataset cloning guidance: PR #27144 by @glenn-jocher clarifies the Ultralytics Platform workflow: visitors use Clone Images, while owners use copy and paste.
  • Platform workflow coverage: PR #27145 by @pderrenger updates guidance for PDF prediction uploads, dense results, model cloning limits, stopped-deployment metrics, and Gemini 4 Carbon.
  • PP-OCRv6 language coverage: PR #27155 by @onuralpszr documents supported languages and characters.
  • Clearer OCR expectations: PR #27157 by @glenn-jocher clarifies which writing systems YOLO OCR can recognize, including that Cyrillic is not yet supported.

:raising_hands: Try It and Share Your Feedback

Install this release with:

pip install --upgrade ultralytics==8.4.211

Browse the full changelog for all changes, then give the update a spin.

What improvements are you seeing? Share your results, hardware, model, and settings in the replies—or let us know if you encounter an issue. Thanks to every contributor who helped make this release possible! :tada: