QwenTeam Announces Qwen-Drive-1.0 Foundation Model for Autonomous Driving

On 2026/09/03, the QwenTeam announced Qwen-Drive-1.0, a vision-language foundation model for autonomous driving that unifies 3D perception and visual question answering at the pretraining stage. It builds on the natively multimodal Qwen3.5-4B. Without changing the pretrained architecture, the model attaches two external modules. A BEV perception head acts as an explicit, inspectable 3D probe. Meanwhile, a Planning Expert generates future ego trajectories through flow matching.

The BEV perception head jointly performs 3D object detection, semantic occupancy prediction, and BEV map segmentation. The Planning Expert is a diffusion transformer tailored to VLM representations. It generates 5-second ego trajectories at 10 Hz, using an optional textual planning reason as a condition.

A staged training recipe combines driving data with general-purpose vision-language supervision to support domain adaptation while mitigating catastrophic forgetting. The training is based on publicly available data comprising a total of 2.83 million samples.

In performance testing, Qwen-Drive-1.0-SFT reached a driving QA average of 69.43. On individual benchmarks, the model scored 77.80 on LingoQA, 7.78 on Ego3D RMSE, and 70.70 on WaymoQA Safety. Under the official LingoJudge protocol, Qwen-Drive-1.0-SFT obtains a LingoScore of 79.4. On general benchmarks, it scored 85.53 on MMBench, 75.87 on MMStar, and 72.67 on MMMU.

On the NAVSIM pseudo-closed-loop benchmark, Qwen-Drive-1.0-RL achieved a PDMS score of 90.7, which rose to 91.4 in a leading-of-6 setting. For open-loop motion planning on WOD-E2E, Qwen-Drive-1.0-RL recorded a 5s val ADE of 1.27 and a test ADE of 2.67.