Alibaba Open-Sources Qwen-Drive-1.0-4B Vision-Language Model for Autonomous Driving
Alibaba's Qwen team has open-sourced Qwen-Drive-1.0-4B, a 4-billion-parameter vision-language foundation model designed specifically for autonomous driving. Built on Qwen3.5-4B, it unifies 3D perception, visual question answering, and motion planning within a single VLM architecture. It is the first vision-language foundation model tailored for autonomous driving while retaining general visual understanding and instruction-following capabilities. By releasing weights and code publicly on GitHub and Hugging Face, Alibaba lowers the barrier for researching end-to-end AI in self-driving technology. Qwen-Drive-1.0 uses a phased training pipeline combining driving supervision with general VLM data, offering both Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL) planning experts. While the RL-tuned model introduces a minor open-loop displacement error, it significantly improves closed-loop safety and alignment with human driver preferences.
## BACKGROUND
Autonomous driving systems traditionally separate perception, prediction, and control into independent pipelines. Evaluating these systems involves open-loop testing, which checks trajectory accuracy against recorded human driving data, and closed-loop or pseudo-closed-loop simulation, which tests how the model continuously interacts with a dynamic reactive environment.