Qwen-Drive-1.0-4B: alibaba's open-source vision-language model for autonomous driving

On September 6, 2026, Alibaba's Qwen team announced the open-sourcing of Qwen-Drive-1.0-4B, a vision-language model tailored for autonomous driving. This model is built on Qwen3.5-4B and is the first of its kind to integrate 3D perception and visual question answering during pre-training, extending further into motion planning while preserving the original architecture of pre-trained vision-language models. The model employs a staged training approach, combining driving supervision with general visual-language data to retain both domain-specific capabilities and general visual understanding. It offers two planning experts—SFT and RL—and has been evaluated across 3D perception, scene understanding, and motion planning. The model uses only publicly available data, standardizing trajectory formats and assessing performance from open-loop to closed-loop scenarios. After reinforcement learning, it achieves significant improvements in human alignment and closed-loop safety with minimal open-loop displacement error. Qwen-Drive-1.0-4B is now available on GitHub, Hugging Face, and ModelScope.