Ultralytics v8.4.188: Smoother Android Deployment and Leaner Depth Exports
Ultralytics YOLO v8.4.188 is here! This release improves Android QNN model loading, removes an unnecessary operation from non-LiteRT depth exports, and refreshes the LiteRT export toolchain. These are export and deployment improvements—model architectures and weights are unchanged.
Android QNN Fixes and Depth Export Improvements
The changes in PR #26872 by @glenn-jocher address two deployment pain points:
- More compatible Android QNN exports: Input and output quantization now stay inside the QNN context, avoiding an
io_name_overridesattribute that newer export tools can add but Android’s built-in QNN Execution Provider cannot read. This should resolve loading issues for affected models, including in apps using the Ultralytics Flutter plugin. - Leaner depth exports outside LiteRT: The export-only ReLU is now limited to LiteRT, where it remains a workaround for an Intel NPU compiler issue. QNN and other export formats avoid that extra operation, potentially improving runtime efficiency.
Updated LiteRT Toolchain and Export Docker Image
The updates in PR #26869 by @glenn-jocher bring the LiteRT export dependencies in line with current mobile NPU runtimes:
litert-torch >= 0.10.0ai-edge-litert >= 2.3.0ai-edge-quantizer >= 0.10.0
The export Docker image now installs the full export dependency set, including LiteRT support. These updates can improve export compatibility for supported devices.
Try It and Share Your Feedback
Upgrade to the release with:
pip install --upgrade ultralytics==8.4.188
Check the v8.4.188 release notes for the release details, or browse the full changelog to review the changes.
Working with Android QNN, LiteRT, or depth models? Give this release a try and let us know how it performs. Share your device, export format, and results below—your feedback helps improve deployment for everyone! ![]()