Qwen releases Qwen-Drive-1.0 driving model
Qwen's Qwen-Drive-1.0 adds 3D perception and motion planning to its general-purpose Qwen3.5-4B model without changing the model's architecture, and it scores higher than larger driving specialists on driving question-answering benchmarks.
The Qwen team has introduced Qwen-Drive-1.0, which it calls the first vision-language foundation model for autonomous driving to combine 3D perception and visual question answering during pretraining. The model also handles motion planning. The post links to a paper, GitHub, Hugging Face and ModelScope.
The model is built on the natively multimodal Qwen3.5-4B, and the pretrained architecture is left unchanged. Two external modules read from the shared vision pathway. The first is a bird's-eye-view (BEV) perception head that handles 3D object detection, semantic occupancy prediction and BEV map segmentation together. The team describes it as an "explicit, inspectable 3D probe." The second is a Planning Expert, a diffusion transformer that uses flow matching to generate 5-second ego trajectories. It can optionally be conditioned on a written planning reason. Training happens in stages and mixes driving data with general vision-language data to limit catastrophic forgetting.
According to Qwen, the SFT variant averages 69.43 on driving QA, ahead of general VLMs and driving specialists including MiMo-Embodied-7B, Cosmos-Reason2-8B and Alpamayo-1.5-10B. Examples include 77.80 on LingoQA, 74.47 on WaymoQA All and 66.13 on SURDS, up from 52.95 for the base Qwen3.5-4B. General benchmarks stay close to the base model: 75.87 on MMStar versus 75.33, and 72.67 on MMMU versus 73.44.
Qwen scored LingoQA with Qwen-Plus as the judge in place of the official LingoJudge, which it said scores leniently and inconsistently. Under the official protocol, the model scores 79.4. Qwen describes its open-loop, pseudo-closed-loop and closed-loop planning results as "highly competitive," but the provided text ends before those numbers appear.