Qwen-Drive-1.0: An Initial Step towards a Vision-Language Foundation Model for Autonomous Driving

Qwen-Drive-1.0 is a unified autonomous-driving vision-language model built on Qwen3.5-4B without changing its pretrained architecture. It adds an external bird’s-eye-view perception head for 3D object detection, semantic occupancy, and map segmentation, plus a Planning Expert that uses flow matching to generate 50 future waypoints over 5 seconds. A four-stage recipe combines driving supervision, general vision-language data, and reinforcement learning, with a 1.54M-example Stage 2 mixture designed to improve driving competence while limiting catastrophic forgetting. On nuScenes, the model reaches 43.95 mAP for 3D detection and 60.99 mIoU for occupancy, while its driving VQA average rises to 69.43, outperforming the base Qwen3.5-4B and competing with larger systems such as NVIDIA’s Cosmos models, Gemma4, and Alpamayo. Reinforcement learning raises the NAVSIM Predictive Driver Model Score to 90.7 and produces a 7.91 Rater Feedback Score on the held-out Waymo Open Dataset end-to-end test split. In closed-loop AlpaSim evaluation, reinforcement learning reduces the off-road rate from 24.0% to 12.0%, although progress also falls, highlighting a safety-versus-aggressiveness trade-off. The central contribution is an inspectable 3D interface and trajectory-generation module that extend a general-purpose VLM while largely preserving broad visual reasoning and instruction following.