Title: Ultralytics v8.3.188 – Faster downloads, smoother dev workflows, clearer outputs
Summary
Ultralytics v8.3.188 brings faster and more reliable downloads, improved export compatibility, clearer JSON outputs, and smoother documentation workflows. It’s a focused release that streamlines setup, development, and benchmarking.
New Features
- Faster, more reliable downloads
- PR #21850 by @glenn-jocher
- Single-request flow removes slow HEAD requests, with smarter disk-space checks (fail-fast on low space), adaptive streaming buffers, precise progress, and modernized type hints/signatures.
Improvements
- Export compatibility
- PR #21758 by @Y-T-G
- Removed ONNX upper bound on Linux/Windows; macOS keeps <1.18.0 to avoid TensorFlow hangs. Updated simplification stack to onnxslim>=0.1.65.
- JSON outputs now include file_name
- PR #21837 by @olena-hul-dataspan
- Adds “file_name” across RT-DETR, YOLO Detect, and YOLO OBB validators for easier traceability and dataset alignment.
- TQDM robustness and developer experience
- PR #21849 by @Laughing-q
- Handles total=0 gracefully and improves type hints.
- Docs developer UX upgrades
- PR #21842 by @glenn-jocher, PR #21848 by @Laughing-q
- Auto-serve docs after build on macOS/Linux with unified logging.
- Benchmark display polish
- PR #21820 by @onuralpszr
- Clear, full, indexed tables with safe null handling.
- Docs/content touch-ups
- PR #21835 by @Laughing-q (typo fixes), PR #21841 by @glenn-jocher (authors metadata)
Bug Fixes
- TensorFlow export stability with newer ONNX on Linux/Windows, while keeping macOS constraints for reliability (PR #21758 by @Y-T-G).
- Progress bars no longer break on empty sequences and reflect progress more accurately (PR #21849 by @Laughing-q).
Purpose and Impact
- Speed and reliability: Faster model and asset downloads with adaptive streaming and early disk checks.
- Fewer dependency conflicts: Relaxed ONNX constraints reduce install friction on Linux/Windows; macOS stays stable.
- Easier result tracking: “file_name” in predictions.json simplifies mapping outputs to source images.
- Better progress feedback: More accurate, robust TQDM behavior.
- Smoother docs workflows: Auto-serve docs and consistent logging for faster iteration.
- Cleaner benchmarks: Readable, comprehensive tables for quick comparisons.
Getting Started
Upgrade with:
pip install -U ultralytics
What’s Changed (PRs and authors)
- Update typo in sam-2.md — PR #21835 by @Laughing-q
- Update mkdocs_github_authors.yaml — PR #21841 by @glenn-jocher
- Serve Docs automatically during development — PR #21842 by @glenn-jocher
- Add downloads.py buffer size — PR #21843 by @glenn-jocher
- Enhance Polars configuration for better benchmark display — PR #21820 by @onuralpszr
- Add file_name to predictions.json — PR #21837 by @olena-hul-dataspan
- Fix TensorFlow export with onnx>=1.18.0 — PR #21758 by @Y-T-G
- Support serving docs automatically on Linux — PR #21848 by @Laughing-q
- Fix TQDM progress bars with empty sequences — PR #21849 by @Laughing-q
- Faster downloads via a single request — PR #21850 by @glenn-jocher
New Contributors
- @olena-hul-dataspan made their first contribution in PR #21837
Try It and Share Feedback
Please upgrade, kick the tires, and let us know how it goes—especially around downloads, exports, and predictions.json. Your feedback helps the community and the Ultralytics team continue improving YOLO.
Links
- Release: v8.3.188
- Full changelog: compare v8.3.187…v8.3.188
Thanks to all contributors and the YOLO community for making each release better than the last!