New Release: Ultralytics v8.3.65 
We’re excited to announce the release of Ultralytics v8.3.65, packed with new features, enhancements, and fixes designed to expand capabilities and improve performance. Here’s a quick overview of what’s new:
Summary
With this release, YOLO models now support export to Rockchip’s RKNN format, enabling seamless AI deployment on Rockchip NPUs like the RK3588 and RK3566. We’ve also included key stability improvements, modernized parts of the codebase, and ensured compatibility with more platforms to enhance developer experience.
Key Features and Improvements
1. Rockchip RKNN Integration
- Export YOLO models to the RKNN format for optimized use on Rockchip hardware.
- New tools and documentation simplify deployment workflows for Rockchip devices.
- Hardware inference is supported via
rknn-toolkit2
, with device compatibility checks included.
2. Stability and Performance Enhancements
- Improved data loader robustness: Safely handles edge-case worker cleanup to reduce crashes.
- CI support for macOS 15, ensuring seamless compatibility with the latest macOS environments.
3. Compatibility Updates
- Dynamic management of
numpy
dependencies for NVIDIA Jetson devices (e.g., Jetson Nano), streamlining TensorRT integration for better reliability.
4. Codebase Modernization
- Replaced mutable Python
set
with immutablefrozenset
throughout the codebase, reducing potential bugs in multi-threaded environments and ensuring data integrity.
5. Documentation Enhancements
- Cleaned and maintained regex for consistent link conversion, improving usability and simplifying updates.
Purpose & Impact
Purpose
- Expand Ultralytics YOLO deployment capabilities to Rockchip’s embedded hardware.
- Enhance stability across modules by addressing error-prone edge cases.
- Optimize the toolset for faster, safer, and more reliable performance in real-world use cases.
Impact
RKNN Support: Easy export of YOLO models to Rockchip NPUs brings high-performance AI to embedded systems.
Stability: Improved robustness ensures fewer crashes and smoother user experiences.
Performance Optimization: More reliable operations for developers working in multi-threaded or embedded environments.
Documentation Quality: Clearer and more user-friendly resources to support new and existing features.
What’s Changed
Here’s a full breakdown of the latest changes:
- Catch and ignore exceptions in data loader cleanup by @Y-T-G:
PR #18772 - Pin
numpy
1.23.5 for JetPack 4 (NVIDIA Jetson Nano) by @lakshanthad:
PR #18783 - Use
frozenset()
for immutability by @glenn-jocher:
PR #18785 - Adopt
macos-15
runners for CI workflows by @glenn-jocher:
PR #18763 - Update regex to convert plaintext links to HTML for docs by @glenn-jocher:
PR #18786 - Add RKNN export support for YOLO models by @IvorZhu331:
PR #16308
Useful Links
- Release Notes: Ultralytics v8.3.65 Release
- Full Changelog:
Compare v8.3.64…v8.3.65
Let Us Hear from You!
We invite you to try out the new features and provide your valuable feedback. Your insights are vital for the continuous improvement of our tools. Explore these updates now, and let us know what you think!
Thank you for being part of the Ultralytics community Let’s keep making AI deployment smarter and faster together!