Alibaba's Qwen open-sources autonomous driving model Qwen-Drive-1.0-4B

Alibaba's Qwen team open-sourced Qwen-Drive-1.0-4B this week, an autonomous driving vision-language model built on Qwen3.5-4B. It's the first vision-language foundation model for self-driving. The model unifies 3D perception and visual question answering during pre-training, and extends to motion planning while keeping the original architecture.

The model uses a staged training approach, combining driving supervision with general vision-language data. It keeps general understanding while gaining driving skills. Qwen-Drive-1.0-4B offers two planning experts: SFT and RL. Tests cover 3D perception, scene understanding, and various motion planning tasks.

According to the team, the model only uses public data to build planning samples and unifies trajectory formats across datasets. The SFT model is competitive on benchmarks. After reinforcement learning, it greatly improves human preference alignment and closed-loop safety, with only a tiny increase in open-loop displacement error.