Qwen-Drive-1.0, Vision-Language Foundation Model for Autonomous Driving Released

The Qwen team released Qwen-Drive-1.0, the first vision-language foundation model designed for autonomous driving. This model integrates 3D perception, visual question answering (VQA), and motion planning by attaching external modules without modifying the pre-trained VLM architecture.

Core Architecture and Features

• Base Model: Based on the native multimodal Qwen3.5-4B, maintaining the pre-trained VLM structure as is.
• BEV Perception Head: Acts as an explicit and inspectable 3D probe that processes multi-view inputs to perform 3D object detection, semantic occupancy prediction, and BEV map segmentation.
• Planning Expert: A diffusion transformer aligned with VLM representations, generating 5-second ego trajectories via flow matching.

Performance and Results

Qwen-Drive-1.0-SFT achieved an average score of 69.43 on autonomous driving VQA, surpassing general-purpose VLMs and autonomous driving-specific models. It also demonstrated competitive performance on major benchmarks such as LingoQA and WaymoQA, effectively adapting to the autonomous driving domain while maintaining general vision-language capabilities (MMBench, MMMU, etc.).