YOLO-Master

What YOLO-Master introduces

YOLO-Master is a YOLO style framework tailored for real time object detection. The paper behind it was accepted by CVPR 2026, and the authors are from Tencent Youtu Lab and Singapore Management University. The README states it marks the first deep integration of mixture of experts, or MoE, into the YOLO architecture for general datasets. The method uses efficient sparse MoE, called ES-MoE, together with lightweight dynamic routing. This gives the model instance conditional adaptive computation, a compute on demand paradigm where the model allocates floating point operations based on scene complexity. The claim is a better Pareto frontier between high precision and ultra low latency. The dynamic routing network guides expert specialization during training and activates only the most relevant experts during inference, which reduces redundant computation while raising detection performance. The README frames the intuition as a shift from static dense computation to input adaptive compute allocation, with more pronounced gains in dense or challenging scenes. A complete end to end pipeline is provided, covering installation, validation, training, inference, and deployment to formats such as ONNX and TensorRT. The project also references advanced utilities as part of its continuous engineering evolution. For practitioners, the appeal is keeping YOLO's speed while adding a routing mechanism that spends computation where the image needs it rather than uniformly across every frame.

Reported results and pipeline

The README reports concrete numbers from MS COCO. YOLO-Master-N reaches 42.4 percent AP at 1.62 milliseconds latency, and the project states this outperforms YOLOv13-N with a 0.8 percent mAP gain while being 17.8 percent faster. Those figures are the headline evidence for the accuracy versus latency claim, and they are tied to a specific model size, the Nano variant. The framework also ships a Model Zoo, referenced through a release tag, and a Hugging Face Spaces demo plus a Colab notebook linked from the badges. Two arXiv identifiers are listed, one for the main paper and one for a YOLO-PEFT paper, and the project builds on the Ultralytics YOLO ecosystem, which the README badges acknowledge. The out of the box pipeline is a selling point: installation, validation, training, inference, and deployment are all provided, with export targets including ONNX and TensorRT for production use. The README mentions continuous engineering evolution that includes advanced utilities, suggesting the tooling around training and export keeps growing. For someone evaluating the framework, the practical path is to start from the Model Zoo or the Colab demo, then follow the installation and validation steps before training a custom model. The AGPL-3.0 license in the repository metadata is important for adoption decisions, since it carries copyleft obligations for networked use.

Where to find it

The project is hosted at the Tencent YOLO-Master GitHub repository, and the homepage links to a Tencent GitHub pages site for the project. The README carries badges for Hugging Face Spaces, Colab, two arXiv papers, the CVPR 2026 acceptance, the Model Zoo release, the AGPL-3.0 license, and Ultralytics. The authors are named in the README with equal contribution noted for two of them, and the affiliations are Tencent Youtu Lab and Singapore Management University. The CVPR 2026 badge indicates peer review acceptance, which is a meaningful signal for a research framework. The Model Zoo release tag, YOLO-Master-v26.02, is where pretrained weights are published according to the badge. Because the framework is built on Ultralytics conventions, users familiar with that ecosystem will recognise the training and inference patterns. The demo on Hugging Face Spaces lets a visitor try detection without local setup, and the Colab notebook gives a runnable starting point. The license is AGPL-3.0, so any network served modification must be shared under the same terms. The source, issues, and releases are on the Tencent YOLO-Master repository, and the documentation and demo links are gathered in the README header. For a team considering adoption, the key facts are the CVPR 2026 paper, the MS COCO numbers for the Nano model, the ONNX and TensorRT export path, and the AGPL-3.0 license.

Editorial conclusion

The project is published under the AGPL-3.0 license and written in Python, with its source at the Tencent YOLO-Master repository.