Ultralytics v8.3.209 — Reliable RT‑DETR → TensorFlow/TFLite exports, tuner fix, and docs/CI updates 
Summary
Ultralytics v8.3.209 focuses on export reliability and tooling polish. RT-DETR exports to TensorFlow/TFLite are now robust by automatically enforcing ONNX opset 19, hyperparameter tuning gets a correctness fix when resuming, and we’ve refreshed CI and documentation assets. YOLO11 and YOLO26 models and other export paths are unaffected by these changes.
New Features
- No new features this cycle—this release concentrates on reliability and developer experience improvements.
Improvements
- More reliable RT-DETR → TensorFlow/TFLite export
- Thanks to @Y-T-G for the RT-DETR TFLite export fix with ONNX opset constraints, which automatically enforces
opset=19during the ONNX step to prevent conversion failures and directly addresses issue #18055. ONNX simplification remains enabled to improve stability. Scope is limited to the TensorFlow SavedModel export path; YOLO11/YOLO26 and other models/exports are unaffected. - Export pipeline dependencies now include
onnxslim>=0.1.71to enhance graph simplification for both ONNX and TensorFlow SavedModel paths.
- Thanks to @Y-T-G for the RT-DETR TFLite export fix with ONNX opset constraints, which automatically enforces
- CI and Docs updates
- We’ve expanded CI visibility by adding badges to more repos in the CI Help Docs, including Ultralytics Stars and CLIP. See the contribution by @glenn-jocher in Add ultralytics/stars to Docs CI page.
- Documentation visuals have been refreshed with a reliable PNG banner and a “Shenzhen” label for YOLO Vision 2025. Props to @sergiuwaxmann for the initial YOLO Vision 2025 Shenzhen Docs banner and to @glenn-jocher for the banner update.
- GPU CI jobs are temporarily paused to reduce flaky runs due to runner availability.
Bug Fixes
- Hyperparameter tuning reliability
- Resuming tuning now correctly tracks and saves the best iteration, thanks to @Y-T-G in the best fitness iteration fix.
Purpose & Impact
- Stronger export reliability for RT-DETR: enforcing ONNX opset 19 avoids the conversion failures many saw with higher opsets and can override a user-provided higher opset to ensure success.
- Smoother export experience: newer
onnxslimimproves ONNX graph simplification, reducing export friction. - More accurate experiments: the tuner fix prevents incorrect “best run” selection when resuming.
- Clearer docs and steadier CI: improved branding and centralized CI visibility, with a temporary GPU CI pause to stabilize pipelines.
Quick Start
- Export RT-DETR to TensorFlow or TFLite from the CLI:
yolo export model=rtdetr.pt format=tfyolo export model=rtdetr.pt format=tflite
- Export from Python:
from ultralytics import YOLO model = YOLO("rtdetr.pt") model.export(format="tf") # TensorFlow SavedModel model.export(format="tflite") # TFLite - If you’re new to RT-DETR, you can skim the RT-DETR model docs for task support and usage notes. For most new projects we recommend starting with YOLO11; this release does not alter YOLO11 or the in‑progress YOLO26 R&D line.
What’s Changed (PRs and Authors)
- Documentation banner for YOLO Vision 2025 Shenzhen by @sergiuwaxmann in PR #22313
- CI Docs page update adding ultralytics/stars by @glenn-jocher in PR #22331
- YOLO Vision 2025 Shenzhen banner update by @glenn-jocher in PR #22332
- Tuner best-iteration fix when resuming by @Y-T-G in PR #22329
- RT-DETR TFLite export fix with opset constraints by @Y-T-G in PR #22314
Get the Release
- You can read the detailed notes on the Ultralytics v8.3.209 release page and browse the full changelog diff from v8.3.208 to v8.3.209.
Try It and Share Feedback
- Upgrade with
pip install -U ultralytics, try the improved exports, and let us know how it goes. We welcome your thoughts in GitHub Discussions and issues, and we appreciate reproducible examples if you encounter any edge cases. Your feedback helps the whole YOLO community move forward.