Researchers at Hugging Face have unveiled Qwen-Drive-1.0, a vision-language foundation model designed for autonomous driving. The model integrates 3D perception, visual question answering, and motion planning into a unified framework, built on the natively multimodal Qwen3.5-4B architecture.
A key feature is the external bird's-eye-view (BEV) perception head, which jointly performs 3D object detection, semantic occupancy prediction, and BEV map segmentation. This head acts as an explicit, inspectable probe of the 3D information available from shared representations, providing a transparent interface to the scene structure.
Additionally, a Planning Expert module conditions on the shared VLM representations to generate future ego trajectories using flow matching. The model employs a staged training recipe that combines driving supervision with general-purpose vision-language data, aiming to acquire driving-specific competence while preserving broad visual understanding and instruction-following abilities.
Experiments demonstrate strong performance in 3D perception and driving scene understanding, with competitive motion-planning results across open-loop, pseudo-closed-loop, and closed-loop settings. The approach notably keeps the pretrained VLM architecture entirely untouched, offering a new-generation base for driving-scenario adaptation.