New Release: Ultralytics v8.4.65

Discussion
, ,
1

Ultralytics v8.4.65 is out! :package::high_voltage:

Summary: Ultralytics v8.4.65 is a deployment-focused release with major wins for mobile inference, especially on Snapdragon NPUs :mobile_phone:. The headline improvements are faster QNN exports, much more efficient semantic segmentation outputs for QNN and CoreML, and a more robust export pipeline overall.

If you deploy Ultralytics YOLO on edge or mobile devices, this is a great update to try.

:glowing_star: Highlights

:rocket: Faster QNN exports for Snapdragon devices

The biggest upgrade in this release comes from PR #24790 by @glenn-jocher.

QNN exports now use channel-last (NHWC) input, which better matches Qualcomm Hexagon NPU execution and camera-native image layouts. That means less input reordering, lower runtime overhead, and better real-world efficiency for mobile apps.

This was implemented with cleaner export wrappers such as QNNModel, replacing fragile post-export graph editing and making the export path easier to maintain.

:brain: Faster semantic segmentation on-device

Also in PR #24790 by @glenn-jocher, semantic segmentation exports for QNN and CoreML now move ArgMax / class-map generation directly into the model graph.

Instead of returning large float logits that apps must decode afterward, exported models can now return a compact per-pixel class map directly. This reduces app-side CPU work and makes on-device segmentation performance much more stable and responsive.

:mobile_phone: Better mobile runtime compatibility

Prediction and validation handling for semantic models was updated so Ultralytics can correctly work with both:

  • traditional logits outputs
  • new exported class-map outputs

This helps the new faster export behavior integrate more cleanly across deployment and evaluation workflows.

:hammer_and_wrench: Improvements

:white_check_mark: RKNN export reliability improvements

PR #24762 by @Laughing-q improves RKNN export support by:

  • enabling FP16 intermediate ONNX export
  • adding RKNN INT8 export test coverage

This should make RKNN deployment workflows more reliable across hardware targets.

:spouting_whale: Docker build improvements

A couple of changes make container workflows cleaner and more reliable:

  • PR #24783 by @glenn-jocher removes an extra Docker workflow assistant trigger
  • Docker install patterns and cleanup behavior were also improved to reduce disk pressure during builds

:books: Better docs and export guidance

This release includes several documentation refreshes and clarifications:

These updates also expand guidance around CoreML, QNN, YOLO26 end-to-end export outputs, and tracking support for OBB models.

:bug: Fixes

:wrench: CI workflow cleanup

PR #24557 by @lakshanthad improves CI behavior by explicitly skipping CUDA training tests on Jetson devices.

This makes automated testing cleaner and easier to maintain.

:bullseye: Why this release matters

Ultralytics v8.4.65 is especially valuable for:

  • Mobile and edge developers deploying to Snapdragon devices
  • Semantic segmentation users who want faster and more stable on-device outputs
  • Teams shipping production inference pipelines that benefit from stronger export reliability
  • General users who want cleaner docs, CI improvements, and smoother deployment workflows

The biggest practical benefit is simple: less postprocessing outside the model, less runtime overhead, and better mobile performance :high_voltage:

:play_button: Try it out

Upgrade with:

pip install -U ultralytics

You can explore the latest models on the Ultralytics Platform, and for new projects we recommend starting with YOLO26, our latest stable model family. YOLO11 remains fully supported as well.

For the complete release details, see the v8.4.65 release page and the full changelog.

:speech_balloon: Feedback welcome

Please give v8.4.65 a try and let us know how it performs in your mobile, edge, and export workflows. Feedback, benchmarks, and bug reports from the community help us keep improving YOLO for everyone :raising_hands: