New Release: Ultralytics v8.3.201

Discussion
, ,
1

Ultralytics v8.3.201 — smoother TensorFlow exports, steadier Docker builds, clearer HUB CLI help :rocket:

Quick summary: This release focuses on export reliability and developer experience. Expect smoother TensorFlow exports, more robust Docker-based exports, better yolo -h help for Ultralytics HUB commands, a small dataset fix, and clearer docs to help you customize YOLO11 with confidence.

YOLO11 remains our recommended default for all use cases.

Highlights

  • TensorFlow export compatibility expanded for non-macOS platforms :wrench:
  • Export robustness improvements and Docker reliability :spouting_whale:
  • Clearer yolo -h context-aware help for Ultralytics HUB :compass:
  • Multimodal training flexibility for YOLOE workflows :brain:
  • SKU-110K dataset CSV parsing fix :package:
  • New Model YAML guide and updated augmentation reference :books:

Explore the release details in the annotated notes below, and view the release page in the GitHub UI via the link to the tag on the release page for v8.3.201.

New Features & Improvements

TensorFlow exports (primary change)

  • Expanded compatibility for ai-edge-litert by relaxing the requirement to >=1.2.0 on Linux/Windows while keeping macOS pinned below 1.4.0 to avoid known issues. See the PR in the discussion about expanding TensorFlow ai-edge-litert support for non-macOS by Glenn Jocher.

  • check_requirements() now accepts tuple[str] alongside Path | str | list[str] for marginally cleaner typing.

Export robustness and Docker reliability

  • The exporter now asserts non-zero file size and logs tuple-returning exporters correctly for accurate success reporting. Details are in the PR for asserting non-zero filesize in Exporter by Glenn Jocher.

  • Docker export images now install TensorRT on-demand to match runtime CUDA, reducing version conflicts. Review the change to use AutoInstall for TensorRT in the PR by Y‑T‑G.

  • Dockerfiles export using a local temp path (/tmp/yolo11n.pt) for EdgeTPU/NCNN/IMX steps to improve stability and reproducibility. See the PR to export in Dockerfile /tmp dir by Glenn Jocher.

Better CLI help for Ultralytics HUB

  • yolo -h now shows context-aware help for HUB subcommands by adjusting help precedence. Check the PR that improves the help message with yolo -h by Y‑T‑G.

Training flexibility for multimodal (YOLOE) workflows

  • Removed a training-mode assertion to allow get_text_pe() during training, making multimodal setups easier. See the PR that removes the training mode assertion in get_text_pe() by Y‑T‑G.

Dataset reliability

  • Fixed CSV parsing for SKU-110K where no-header CSVs previously misaligned columns. Review the PR that fixes the CSV read error in SKU‑110K.yaml by Y‑T‑G.

Documentation improvements

  • Added a comprehensive reference on configuring and scaling models, using TorchVision backbones, and integrating custom modules. Read the new Model YAML Configuration Guide and see the docs PR by Y‑T‑G for more context.

  • Augmentation docs now include a “Supported Tasks” column that clarifies applicability across detect/segment/pose/OBB/classify. See the PR that adds the Supported Tasks column by Y‑T‑G.

Bug Fixes

  • Exporter now detects empty outputs early and logs correct artifact paths and sizes :white_check_mark:
  • SKU-110K CSV parser fixed to handle no-header CSVs correctly :white_check_mark:

Try it

Upgrade and try a TensorFlow SavedModel export with the updated dependency handling:

pip install -U ultralytics
yolo export model=yolo11n.pt format=saved_model

For HUB-related commands, get context-aware help directly:

yolo -h
yolo hub -h

What’s Changed (PRs and authors)

  • Fix CSV read error in SKU‑110K.yaml download script, contributed by Y‑T‑G in the PR fixing the CSV read error in SKU‑110K.yaml.
  • Remove training mode assertion in get_text_pe(), contributed by Y‑T‑G in the PR removing the training mode assertion in get_text_pe().
  • Update Dockerfile‑export to use AutoInstall for TensorRT package, contributed by Y‑T‑G in the PR to use AutoInstall for TensorRT package.
  • Export in Dockerfile /tmp dir, contributed by Glenn Jocher in the PR to export in Dockerfile /tmp dir.
  • Display help message with yolo -h, contributed by Y‑T‑G in the PR to display the help message with yolo -h.
  • Assert non-zero filesize in Exporter, contributed by Glenn Jocher in the PR asserting non-zero filesize in Exporter.
  • Add “Supported Tasks” column to augmentation arguments table, contributed by Y‑T‑G in the PR adding the Supported Tasks column.
  • Docs: Add comprehensive model YAML configuration guide, contributed by Y‑T‑G in the PR adding the comprehensive model YAML configuration guide.
  • Expand TensorFlow ai-edge-litert≥1.2 support (non‑macOS), contributed by Glenn Jocher in the PR expanding TensorFlow ai-edge-litert support for non-macOS.

You can browse every diff in the comparison view that shows the full changelog from v8.3.200 to v8.3.201.

Notes on models

  • YOLO11 is the latest stable and recommended model for all use cases.
  • YOLO12 and YOLO13 are not recommended due to training instability, higher memory/latency costs, or non-reproducible benchmarks compared to YOLO11.

Feedback welcome

We’d love your feedback and bug reports to keep smoothing the edges. Try the new release, share your experience, and let us know what you’d like to see next. Thanks to our amazing community and the Ultralytics team for the continued momentum!