Qwen Releases 'Qwen-Drive-1.0', a Vision-Language Foundation Model for Autonomous Driving

VLM Architecture Specialized for Autonomous Driving The Qwen team and Huazhong University of Science and Technology (HUST) have released Qwen-Drive-1.0, a vision-language foundation model for autonomous driving. Based on the native multimodal model Qwen3.5-4B, this model adopts a structure combining two external modules: the BEV Perception Head and the Planning Expert. The BEV Perception Head handles 3D object detection, semantic occupancy prediction, and BEV map segmentation, while the Planning Expert generates future ego trajectories based on shared VLM representations.

Training Methodology and Performance The model was trained using a staged training approach, jointly learning from autonomous driving-specific supervised data and general vision-language data. In the SFT (Supervised Fine-Tuning) stage, the Planning Expert was trained, followed by the application of reward optimization in the RL (Reinforcement Learning) stage. Performance evaluation results showed that the PDMS score on the NAVSIM v1.1 benchmark improved from 88.2 in SFT to 90.7 with RL applied. In the NVIDIA Open Dataset E2E test, the RFS score also improved from 7.78 to 7.91 with RL applied.

General Capabilities and Spatial Understanding Despite the autonomous driving-specific training, general VQA and reasoning capabilities remained at a level similar to the original Qwen3.5-4B model. Major benchmarks such as MMBench, MMStar, and MMMU showed only minor fluctuations compared to the original model, with some items even showing improved performance. However, performance degradation was observed compared to the original model on certain spatial understanding benchmarks such as EmbSpatial.

Open Source Distribution and Usage The model has been released on GitHub and Hugging Face under the Apache 2.0 license, requiring a GPU with 24GB or more of memory and a Python 3.10 environment. The VLM, SFT/RL-based Planning Expert, and Perception modules are provided as separate files, and scripts/demo.py allows visualization of trajectory predictions and the Reasoning process in various scenarios, such as night intersections and left/right turns.