New Release: Ultralytics v8.4.127

Discussion
, ,
1

:rocket: Ultralytics v8.4.127: More Reliable Exports, Training, and Deployment

Ultralytics v8.4.127 makes exported YOLO models reliably load with the correct task and model family across all 20 supported formats. This release also improves OpenVINO and CoreML deployment, training recovery, result handling, tracking, datasets, and documentation.

Review the complete details on the Ultralytics v8.4.127 release page before exploring the highlights below.

:glowing_star: New Features

  • Correct task detection across all 20 export formats: PR #25886 by @artest08 now reads embedded export metadata instead of guessing from filenames or paths. This preserves segmentation masks and pose keypoints while correctly routing RT-DETR outputs through both the Python API and CLI.
  • CoreML NMS for segmentation and pose: PR #25873 by @SergioAlmeida29 adds nms=True support while keeping masks and keypoints aligned with retained boxes. :red_apple:
  • Direct access to official Platform datasets: PR #25892 by @glenn-jocher links all 52 publicly hosted datasets from the documentation, making it easier to preview, clone, annotate, train, and deploy through the Ultralytics Platform.

:high_voltage: Improvements

  • PR #25893 by @glenn-jocher applies the correct OpenVINO throughput hint for single-device dynamic batching.
  • PR #25881 by @Y-T-G documents YOLO26 Objects365 pretraining, COCO fine-tuning, embedded training logs, training curves, and checkpoint code revisions. :books:
  • PR #25856 by @raimbekovm rewrites the YOLOE documentation and clarifies the evaluation protocol on both LVIS results pages.
  • PR #25871 by @raimbekovm selects the correct YOLOE model and validator by task in the shared trainer.
  • PR #25891 by @Rahulbiradar9 serializes custom augmentations as portable text representations in checkpoints, improving portability and reproducibility.
  • PR #25888 by @Kagura-Ahad adds support for Path filenames in imread.
  • PR #25166 by @Y-T-G fits TQDM output to the available terminal width.
  • PR #25900 by @Y-T-G streams live progress-bar updates to console-log consumers.
  • PR #25884 by @raimbekovm moves documentation media to the CMS assets CDN for more reliable previews in tools such as Slack and Telegram.

:hammer_and_wrench: Bug Fixes

  • Safer training resume: PR #25816 by @Ajaysingh-2003 prevents custom pretrained weights from replacing checkpoint weights during resume, preserving the model, optimizer, scaler, EMA, epoch, and distillation state from last.pt. :counterclockwise_arrows_button:
  • Dense result preservation: PR #25880 by @JESUSROYETH ensures semantic segmentation and depth indexing retains complete dense maps rather than returning only one image row.
  • DeepOCSORT matching: PR #25875 by @JESUSROYETH passes last-observation IoU into OCR appearance fusion, preventing incorrect matches between zero-overlap objects.
  • TrackTrack behavior: PR #25877 by @JESUSROYETH aligns new-track prediction and confirmation with the reference implementation.
  • YOLOE training stability: PR #25867 by @raimbekovm fixes prompt-free segmentation crashes during final validation and re-enables CLIP tests on Python 3.12.
  • OpenVINO INT8 stability: PR #25898 by @glenn-jocher addresses dynamic-shape segmentation faults on Intel AMX CPUs and replaces a dead documentation link.

:waving_hand: Welcome, New Contributors

A warm welcome to @Ajaysingh-2003, who made a first contribution through PR #25816, and @Kagura-Ahad, who joined through PR #25888. Thank you for helping improve Ultralytics YOLO! :tada:

:package: Try v8.4.127

Upgrade to the latest release with:

pip install -U ultralytics

Whether you export specialized models, resume long training runs, deploy with OpenVINO or CoreML, or work with dense prediction tasks, v8.4.127 should provide a more dependable workflow.

Explore every change in the full v8.4.126 to v8.4.127 changelog, try the release, and let the Ultralytics team and YOLO community know how it performs in your projects. :rocket: