New Release: Ultralytics v8.4.61

Discussion
, ,
1

Ultralytics v8.4.61 is out :rocket:

Summary: Ultralytics YOLO v8.4.61 is a stability and export reliability release focused on making deployment smoother and more predictable. The headline fixes are for INT8 ONNX export failures on small calibration datasets and read-only onnx2tf patching, along with stronger CI coverage, better TensorRT compatibility, SAM cleanup improvements, and a broad docs refresh across YOLO26 and the Ultralytics Platform.

If you’re deploying YOLO26 to edge or production environments, this release is especially worth updating to. :white_check_mark:

Highlights

:wrench: More reliable exports

The biggest practical improvement in v8.4.61 is export robustness:

  • INT8 ONNX export fix: exporting could previously fail when the calibration dataset had fewer images than the requested batch size. This is now fixed in PR #24721 by @glenn-jocher, improving reliability for small test sets and static-batch quantized exports.
  • Read-only onnx2tf patching fix: the same PR #24721 by @glenn-jocher also resolves issues in environments where patching files was blocked by read-only permissions.

:test_tube: Stronger CI for deployment formats

We expanded export validation and hardened CI so issues are caught earlier:

:robot: Better compatibility on older TensorRT setups

For users on older Jetson or TensorRT environments, PR #24647 by @yzfzzz now automatically disables end2end export for TensorRT versions older than 8.5.0, avoiding hard-to-debug failures.

:bullseye: SAM post-processing improvement

PR #24688 by @raimbekovm fixes duplicate suppression in remove_small_regions, improving mask cleanup for affected SAM workflows.

:books: Docs and product clarity refresh

This release also includes a substantial round of documentation and correctness updates across YOLO26, YOLOE-26, semantic segmentation messaging, and Platform docs:

Bug Fixes

Why this matters

This release is less about new models and more about making YOLO26 deployment safer, smoother, and more production-ready :nut_and_bolt:

You should see benefits such as:

  • fewer export failures in real deployment workflows
  • better behavior on edge hardware like Jetson, Rockchip, and ARM systems
  • more confidence in ONNX, RKNN, and TensorRT export paths
  • cleaner docs and more accurate warnings across the product

New contributor :raising_hands:

A big welcome to @yzfzzz, who made their first contribution with PR #24647!

Try it out

Update with:

pip install -U ultralytics

Then test your export workflow again, especially if you use INT8 ONNX, TensorRT, or edge deployment targets.

You can explore the full release on the v8.4.61 release page, and review everything included in the full changelog from v8.4.60 to v8.4.61.

As always, give it a try and let us know how it works for your workflows — feedback from the community helps keep YOLO improving :blue_heart: