Exciting New Release: Ultralytics YOLO v8.3.70!
Ultralytics is thrilled to announce the release of YOLO v8.3.70, packed with powerful enhancements for smoother workflows, expanded hardware compatibility, and usability improvements. 
Dive into what’s new below!
Summary
The v8.3.70 update brings:
- Export enhancements with added flexibility for dataset configurations.
- PyTorch 2.6 compatibility for seamless integration with the latest updates.
- Granular benchmarking capabilities for format-specific optimization.
- NVIDIA DLA support, enabling low-latency deployment on specialized hardware.
- Improved
numpystability for consistent performance. - Updated tutorials and better documentation for user onboarding.
Explore the full changelog here.
Key Changes
Sony IMX500 Export Update
- What’s New: Added support for the
dataargument in exports to better control dataset configurations and optimize quantization with models in formats like OpenVINO, TensorRT, and TF Lite. - PR: Add
dataargument to Sony IMX500 export by @lakshanthad.
PyTorch 2.6 Compatibility
- What’s New: Updated Torch-Torchvision mappings for compatibility with PyTorch 2.6.
- PR: Update torchvision compatibility table for
torch 2.6by @glenn-jocher.
Format-Specific Benchmarking
- What’s New: Benchmarks can now target specific export formats like ONNX, allowing deeper insights into performance metrics.
- PR: Add support for single export format benchmarking by @RizwanMunawar.
NVIDIA DLA Core Usage
- What’s New: Support for running YOLO models on NVIDIA DLA cores for optimized inference on hardware-specific platforms.
- PR: Add DLA-specific core usage by @AbelHaro.
Enhanced CI Stability
- What’s New: Pinned
numpy<=2.1.1to resolve compatibility issues during CI tests for OpenVINO and TensorFlow Lite. - PR: Pin
numpyfor CI stability by @lakshanthad.
Documentation Refinements
- What’s New: Added a new tutorial video and refined docs for clarity—check it out to get started faster!
- PRs:
- Add video tutorial to docs by @RizwanMunawar.
- Minor Docs edits by @LexBarou.
Purpose & Impact
- Export Workflows: Empower users with better quantization and deployment control via the new
dataargument during export. Perfect for edge environments. - Torch Compatibility: Stay ahead with PyTorch 2.6’s latest features without breaking workflows.
- Benchmark Smarter: Dive deeper into how models perform in export-specific formats for targeted optimizations.
- Hardware Optimization: Unlock faster inferences using NVIDIA DLA-specific optimizations for specialized hardware.
- Fail-Proof
numpy: Ensure test stability and robustness during development and CI. - User-Friendly Docs: Learn faster with clearer visuals and refined templates.
What’s Changed
- Update torchvision compatibility table for
torch 2.6by @glenn-jocher. - Add YouTube tutorial to docs by @RizwanMunawar.
- Single export format benchmarking by @RizwanMunawar.
- Sony IMX500
datafor export by @lakshanthad. - Add DLA-specific core support by @AbelHaro.
- Pin
numpyversion for CI stability by @lakshanthad. - Minor Documentation tweaks by @LexBarou.
New Contributors
We’re excited to welcome new contributors to the community:
- @LexBarou: Minor Docs edits.
- @AbelHaro: Add DLA specific core usage.
Ready to Try?
Update to v8.3.70 today and take advantage of these improvements:
pip install ultralytics --upgrade
We’d love to hear your feedback—jump into Discussions or report any issues on GitHub.
Thank you for your continued support in building an amazing YOLO community. 