Alibaba's Qwen team has unveiled Qwen-Drive-1.0, a vision-language foundation model designed specifically for autonomous driving. It is the first such model to unify 3D perception and visual question answering (VQA) at the pretraining stage, while also extending to motion planning—all without modifying the underlying pretrained VLM architecture.
Built on the natively multimodal Qwen3.5-4B, Qwen-Drive-1.0 attaches two external modules: a Bird's-Eye-View (BEV) perception head that performs 3D object detection, semantic occupancy prediction, and BEV map segmentation, and a Planning Expert that generates future ego trajectories using flow matching. This design preserves the general capabilities of the VLM while adding specialized driving skills.
Through staged training, the model substantially boosts autonomous driving performance across three key tasks: 3D perception, driving visual question answering, and motion planning. The training recipe unifies cross-dataset labels, rewrites responses, and combines driving data with general-purpose vision-language supervision, mitigating catastrophic forgetting during domain adaptation.
In evaluations, Qwen-Drive-1.0-SFT achieved a driving QA average of 69.43, outperforming both general-purpose VLMs and driving-specific specialists. For motion planning, the model demonstrates strong results in open-loop, pseudo-closed-loop, and closed-loop settings, trained entirely on public datasets totaling 2.83 million samples.
The researchers note that consistency between textual reasoning and generated trajectories remains a challenge, which they plan to address in future work. The project is open-sourced, with code and models available on GitHub and Hugging Face.