Announcing YOLO v8.3.105 
We are thrilled to introduce YOLO v8.3.105, packed with updates to simplify workflows, enhance flexibility, and improve usability. This release focuses on streamlining processes, expanding integration capabilities, and refining visual clarity for object analytics.
Summary
What’s New:
- Removed unused
save_hybridargument for easier validation. - New
deviceparameter enables seamless model exports to GPU, CPU, or MPS. - Improved object counting visuals with adjustable background scaling.
- Added YOLOv8 MNN C++ example for lightweight framework enthusiasts.
- Fixed YOLOE-PF export dimension issues, ensuring better compatibility.
These changes focus on enhancing developer productivity and robustness for diverse use cases, whether in research, development, or deployment.
New Features and Improvements
Simplified Validation
The save_hybrid argument has been removed, making validation workflows easier to manage while preventing potential miscalculations in mAP scores.
Flexible and Optimized Export
The new device parameter allows users to specify the hardware (GPU, CPU, MPS) during export across formats like ONNX, TensorRT, CoreML, and more. This ensures optimized compatibility for various devices, including edge deployments.
Enhanced Object Counting Visualization
We’ve introduced a margin parameter that refines the scaling of text background in object counting analytics, improving real-time display clarity and professional results.
MNN C++ Example
YOLOv8 now integrates seamlessly with the MNN framework, offering developers a ready-to-use example for incorporating YOLOv8 capabilities in lightweight, C+±based architectures.
Reliable YOLOE-PF Exports
A fix for YOLOE-PF model export addresses dimension issues, ensuring accuracy and alignment during deployment.
Changes Overview
What’s Changed:
- Support for adjustable text backgrounds in object counting: By @RizwanMunawar in PR #20073.
- Updated README for MNN example: By @vnk8071 in PR #20065.
- Added
deviceargument to model exports: By @lakshanthad in PR #20009. - Fixed YOLOE-PF export dimensions: By @Y-T-G in PR #20072.
- Removed unused
save_hybridargument: By @Laughing-q in PR #20067.
For a full list of changes, see the v8.3.105 Changelog.
Get Started
This release is now live! You can download YOLO v8.3.105 and start exploring these improvements today.
We warmly invite you to try it out, share your insights, and help us continue improving by providing feedback. Whether you’re using YOLO for real-time detection, segmentation, pose estimation, or deployment, this update is sure to enhance your experience.
Thank you for being part of the Ultralytics community! Together, we are pushing the boundaries of what’s possible in AI and computer vision.
Happy coding! 