Qwen Drive 1.0: A Vision-Language Foundation Model for Autonomous Driving (github.com)

The Qwen Team from Huazhong University of Science and Technology has announced the release of Qwen-Drive 1.0, a pioneering vision-language foundation model designed specifically for autonomous driving applications. This model builds on the architecture of the pretrained Qwen3.5 vision-language model but enhances it by integrating 3D perception, visual question answering (VQA), and motion planning into a cohesive framework. The model features a BEV Perception Head for advanced 3D object detection and segmentation, along with a Planning Expert that generates ego trajectories based on shared representations. The result is a robust system capable of fulfilling both driving-specific and general vision-language tasks. The significance of Qwen-Drive 1.0 lies in its innovative staged training strategy that combines driving-focused supervision with general vision-language data, achieving a nuanced balance of specialized driving competence and broad visual comprehension. With impressive benchmark outcomes—such as improved accuracy in driving scene understanding and causal reasoning—this model positions itself as a substantial advancement in the field of AI for autonomous vehicles. Developers and researchers can access the model on platforms like Hugging Face and ModelScope, provided with detailed documentation for various tasks, ensuring that the AI/ML community can leverage its capabilities effectively.