New Release: Ultralytics v8.4.80

Discussion
, ,
1

Ultralytics v8.4.80 is here :rocket:

Quick summary: Ultralytics v8.4.80 is a deployment-focused release that makes model export precision much simpler with a new unified quantize argument, while also improving multi-GPU validation, OBB training stability, network reliability, and documentation across the Ultralytics ecosystem.

If you export models regularly, validate across multiple GPUs, or train OBB models, this release should make your workflow smoother and more robust. You can explore the release on GitHub Releases.

:glowing_star: Highlights

New unified quantize export argument

The biggest change in v8.4.80 is the new quantize export argument, introduced in PR #24918 by @onuralpszr.

This replaces the older half=True and int8=True export switches with a cleaner, more flexible interface:

  • quantize=32 for FP32
  • quantize=16 for FP16
  • quantize=8 for INT8
  • Advanced modes like quantize="w8a16" where supported

This update makes export settings easier to read, easier to automate, and better aligned with modern hardware-specific quantization workflows. Existing half and int8 arguments remain supported for now, so current scripts should continue working during migration. :white_check_mark:

Example:

from ultralytics import YOLO

model = YOLO("yolo26n.pt")
model.export(format="onnx", quantize=16)

Export docs and integrations updated

To support the new quantize workflow, export examples and docs were refreshed across many backends and deployment targets, helping keep Ultralytics YOLO export guidance consistent and future-ready. :package:

Improvements

Better distributed validation results across GPUs

In multi-GPU validation, confusion matrices are now gathered across all DDP workers instead of reflecting only a single process. This improvement landed in PR #24803 by @Y-T-G, and should produce more accurate confusion matrix plots for distributed evaluation. :chart_increasing:

More stable OBB training

A fix in PR #24933 by @Y-T-G improves training stability for oriented bounding box models by preventing very small sub-stride boxes from triggering invalid loss behavior or NaNs. This is especially helpful for small-object OBB workloads. :bullseye:

Improved request timeout and retry behavior

Network-related reliability was strengthened with:

These changes should reduce cases where requests stall indefinitely and improve failure handling. :globe_with_meridians:

MLflow integration tests restored

PR #24728 by @raimbekovm restores and modernizes MLflow integration tests for newer MLflow versions, improving confidence in experiment logging support. :test_tube:

:books: Documentation updates

Several docs were improved in this release:

  • Tracking persistence and tracker selection were clarified in PR #24889 by @joeydufourd, including better guidance around persist=True, botsort.yaml, and bytetrack.yaml
  • The isolating segmentation guide was fixed in PR #24701 by @raimbekovm
  • Ultralytics Platform dataset upload docs now better explain browser video codec support in PR #24802 by @laodouya, including clearer upload expectations for browser-based video handling :movie_camera:

For users building end-to-end workflows, the Ultralytics Platform documentation remains the best place to learn more about annotating datasets, training, deploying, and monitoring models.

:raising_hands: New contributor

A big welcome to @joeydufourd, who made their first contribution with PR #24889. Thank you for helping improve the docs and user experience! :tada:

Why this release matters

v8.4.80 may not be a major model-launch release, but it delivers meaningful quality-of-life improvements for production and research workflows:

  • Simpler export precision settings with quantize
  • Better support for future mixed-precision deployment
  • More accurate validation outputs in distributed setups
  • More stable OBB training
  • Fewer frustrating request hangs
  • Clearer docs for tracking, segmentation, and Platform uploads

For new projects, we recommend using Ultralytics YOLO26, our latest stable model family, which is smaller, faster, more accurate than YOLO11, and natively end-to-end.

Try it out

Update with:

pip install -U ultralytics

Then give the new export flow a spin with your favorite model, and let us know how it works for you. Feedback, bug reports, and real-world deployment notes are always appreciated. :speech_balloon:

You can also review the complete set of changes in the full changelog.