New Release: Ultralytics v8.4.114

Discussion
, ,
1

:rocket: Ultralytics v8.4.114 Released

Ultralytics v8.4.114 improves reliability across Ultralytics Platform workflows, exported-model validation, edge inference, training pipelines, and advanced vision tasks. Expect clearer errors, simpler validation, faster LiteRT CPU inference, and more dependable deployment workflows. :high_voltage:

:glowing_star: Highlights

  • Clearer Platform errors: Actionable API messages now surface correctly, while invalid credentials and failed console uploads stop unnecessary retries.
  • Simpler exported-model validation: Static exports automatically reuse the image size stored in their metadata.
  • Faster edge inference: LiteRT now uses configured CPU threads for multi-core execution, with updated Raspberry Pi 5 benchmarks for YOLO26n and YOLO26s.
  • More reliable exports: TorchScript, NCNN, and Paddle workflows receive important compatibility fixes.
  • Stronger vision pipelines: Fixes cover SAM3, pose rendering, class activation maps, semantic segmentation, depth training, and distributed validation.

:new_button: New Features and Performance

  • PR #25562 by @lakshanthad enables multi-core Google LiteRT CPU inference using the configured thread count.
  • PR #25564 by @lakshanthad updates Raspberry Pi 5 LiteRT benchmarks for YOLO26n and YOLO26s, reflecting substantially lower latency.
  • PR #25559 by @synml reuses imgsz from export metadata when validating static ONNX, TensorRT, OpenVINO, and similar models. :white_check_mark:

:hammer_and_wrench: Platform and Deployment Improvements

  • PR #25581 by @glenn-jocher surfaces detailed Platform errors, stops requests early when no API key is available, disables further attempts after invalid credentials, and prevents console-output upload failures from creating retry loops.
  • PR #25568 by @Y-T-G fixes NCNN and Paddle export for prompt-free YOLOE models.
  • PR #25569 by @Y-T-G resolves TorchScript GPU inference device mismatches by keeping generated anchors on the runtime device.

:bug: Bug Fixes

  • PR #25563 by @JESUSROYETH defines the required mask_threshold for SAM3 semantic prediction, preventing an AttributeError.
  • PR #25582 by @leoventuroso prevents class_activation_map from raising an IndexError for out-of-range class IDs.
  • PR #25565 by @JESUSROYETH ensures Annotator.kpts() renders keypoints and limbs located exactly on image borders.
  • PR #25567 by @Zenka737 fixes analytics line-chart counts resetting before update_graph consumes them.
  • PR #25576 by @Y-T-G prevents distributed validation failures when the total batch size exceeds the number of validation images.
  • PR #25580 by @raimbekovm applies BGR augmentation correctly during semantic segmentation and depth training.

:books: Documentation and Maintenance

:waving_hand: New Contributor

A warm welcome to @leoventuroso, who made their first contribution in PR #25582. Thank you! :tada:

:package: Try It Today

Upgrade to the latest release with:

pip install -U ultralytics

Visit the Ultralytics v8.4.114 release page for the release details, or review the complete v8.4.113 to v8.4.114 changelog to inspect every change.

Please try the release and share your feedback, benchmarks, and any issues you encounter. Thanks to every contributor and the wider YOLO community for helping make Ultralytics YOLO faster and more reliable! :raising_hands: