New Release: Ultralytics v8.4.161

Discussion
, ,
1

:rocket: Ultralytics v8.4.161 is here

Ultralytics YOLO v8.4.161 focuses on reliability: smoother training resumes, more resilient prediction and plotting, better dataset handling, and clearer export and Ultralytics Platform guidance. See the v8.4.161 release for the release details.

:hammer_and_wrench: Bug fixes

  • Six fixes in one update: RT-DETR tensors move correctly between devices; older Adam/AdamW checkpoints retain their runtime fused setting on resume; empty image files no longer crash prediction; results load from paths containing brackets; plots handle odd numbers of panels; and Platform uploads can use last.pt when save=False leaves no best.pt. See PR #26301 by @glenn-jocher.
  • CoreML export errors are easier to diagnose: removing an invalid .mlmodel fallback prevents it from obscuring the original save error. See PR #26298 by @amanharshx.

:high_voltage: Dataset and training improvements

  • .tgz dataset archives now extract correctly in PR #26294 by @aswanth-07.
  • Semantic datasets skip polygon resampling when no augmentation needs the polygons, reducing unnecessary processing. See PR #26295 by @raimbekovm.
  • CLI fuzz testing now recognizes wrapped missing-dataset-split errors as expected. See PR #26299 by @aswanth-07.

:blue_book: Clearer export and Platform guidance

  • Export documentation now covers the device argument for OpenVINO, Hailo, and Ascend in PR #26286 by @raimbekovm, and explains why static exports can produce different predictions on non-square images in PR #26285 by @Vaishnavi220506.
  • Platform documentation has been refreshed across image copying, face blurring, annotation providers, deployment, monitoring, and related API and billing details. See PR #26297 by @raimbekovm.

:tada: Welcome to first-time contributor @Vaishnavi220506! Thanks to every contributor and the YOLO community for helping make this release possible.

Give v8.4.161 a try and let us know how it works for your training, inference, and export workflows. For every change, browse the full changelog.