Qwen-Drive-1.0

The Qwen Team (with Huazhong University of Science and Technology) extends the Qwen line into autonomous driving: Qwen-Drive-1.0 keeps the natively multimodal Qwen3.5-4B as its shared VLM backbone and unifies 3D perception, driving VQA, and motion planning in one model — the first driving system to the authors' knowledge doing all three in a single pretrained VLM. An external bird's-eye-view perception head jointly performs 3D object detection, semantic occupancy prediction and BEV map segmentation (doubling as an inspectable probe of the 3D structure in the shared representations), and a Planning Expert conditions on the same representations to generate ego trajectories.

A staged recipe mixes driving supervision with general vision-language data so driving competence lands without sacrificing general VQA or instruction following; evaluations span open-loop, pseudo-closed-loop and closed-loop planning. The 4B checkpoint is live on ModelScope; the announced HuggingFace (Qwen/Qwen-Drive-1.0-4B) and GitHub (QwenLM/Qwen-Drive-1.0) listings were not yet public as of 2026-09-02. Joins Alibaba's embodied line (Qwen-Robot, RynnWorld) on the driving side.