New Release: Ultralytics v8.4.212

:rocket: Ultralytics v8.4.212: Expanded OCR and More Reliable Workflows

Ultralytics v8.4.212 expands OCR to more writing systems and improves model discovery, data imports, training reliability, and inference compatibility. This release also brings safer label handling, richer validation exports, and deployment fixes across multiple runtimes.

:globe_showing_europe_africa: New Features and Platform Updates

OCR for more languages and scripts

Choose a PP-OCRv5 recognizer for Korean, Cyrillic, Arabic, Devanagari, Thai, Greek, Tamil, or Telugu text. These models run through ONNX Runtime, without requiring PaddlePaddle or PaddleOCR.

Select the recognizer that matches your expected text: OCR does not automatically detect or switch languages, and Hebrew is not included in these additions.

Easier model discovery and dataset imports

Ultralytics Platform documentation now covers the Explore Models tab, where you can search public models by what they detect, browse results, and try or clone models. Updated guides also explain direct imports from hosted Label Studio projects for detection, oriented-box detection, segmentation, and classification.

  • PR #27160 by @pderrenger documents Explore Models, HEIC previews in Predict, and dataset status in Recent Activity.
  • PR #27163 by @glenn-jocher refreshes VLM leaderboard picks, model deployment and ranking wording, and Predict screenshots.
  • PR #27176 by @glenn-jocher clarifies cloning to another workspace versus duplicating in place with Copy and Paste.
  • PR #27178 by @pderrenger documents direct Label Studio imports and image deletion confirmation.
  • PR #27180 by @glenn-jocher updates Label Studio and Roboflow import screenshots to show the actual storage limit.

:toolbox: Training and Validation Improvements

Training cancellation is more dependable, older macOS MPS configurations receive determinism fixes, and exported validation summaries include more task-specific metrics.

  • PR #27161 by @artest08 fixes distributed training hangs when Platform cancels a run mid-epoch.
  • PR #27159 by @raimbekovm restores deterministic semantic, depth, and segmentation training on MPS with PyTorch versions below 2.3.
  • PR #27162 by @raimbekovm divides the dataloader worker cap by the number of GPUs in the run.
  • PR #27184 by @raimbekovm adds per-class mask mAP, pose mAP, and semantic pixel accuracy to validation CSV and JSON exports.

:high_voltage: Inference and Deployment Fixes

This release improves runtime compatibility and CPU efficiency while addressing several task-specific inference issues.

  • PR #27175 by @Zephyr4772 keeps ONNX inference running on the GPU with CUDA 12 PyTorch installations.
  • PR #27182 by @var1914 improves CPU ONNX prediction speed on hyperthreaded machines and in CPU-limited containers.
  • PR #27185 by @glenn-jocher avoids a redundant 3 GB TensorRT download on JetPack 7, DGX, and NVIDIA containers.
  • PR #27177 by @raimbekovm fixes incorrect MNN detections with device=mps or permuted tensor inputs.
  • PR #27171 by @raimbekovm fixes YOLOE tracking with native ReID features.
  • PR #27181 by @raimbekovm corrects misplaced Edge TPU pose keypoints for models with x, y-only keypoints, such as tiger-pose.

:shield: Safer, More Efficient Data Handling

:books: Documentation and Community

:tada: A special welcome to @AhmetMert01 for their first contribution, and thank you to everyone who helped improve this release!

:rocket: Upgrade and Share Your Feedback

Install the release with:

pip install --upgrade ultralytics==8.4.212

Explore the v8.4.212 release notes for the release overview, or review the full changelog for all changes.

Give v8.4.212 a try and let us know how it works for you! Share your OCR results, deployment experiences, or any issues in the replies below. :speech_balloon: