Alibaba's Qwen team has released Qwen-Drive-1.0-4B, an open-source vision-language model that combines driving-scene understanding with vehicle motion planning1,2. The model was developed in collaboration with Huazhong University of Science and Technology.
Qwen-Drive-1.0-4B uses the pretrained Qwen3.5-4B multimodal backbone as its vision-language foundation, keeping that base architecture unchanged. On top of it, the release attaches modules for bird's-eye-view (BEV) 3D perception and ego trajectory generation. The BEV perception head jointly performs 3D object detection, semantic occupancy prediction, and BEV map segmentation. Qwen describes that head as an explicit, inspectable 3D probe of the shared representations rather than a separate specialist network that replaces the VLM or hides intermediate geometry from developers.
The shared vision-language model retains its ability to answer free-form questions about driving scenes. General knowledge and spatial benchmarks remain broadly in line with the base Qwen3.5-4B, according to the project card and technical report materials.
Motion planning is handled by a Planning Expert that conditions on VLM features and produces future ego trajectories through flow matching. Two planner checkpoints ship in the package. The planner-sft variant is imitation-trained and supports both direct and reasoning-style planning. The planner-rl variant is further reward-optimized on measures such as NAVSIM PDMS and Waymo Open Dataset end-to-end scores and is recommended for reasoning planning mode. A staged training recipe mixes driving supervision with general-purpose vision-language data, intended to gain domain competence while limiting catastrophic forgetting of broad visual skills.
Released evaluation tables show the 4B-class system competing on open-loop prediction, pseudo-closed-loop, and closed-loop evaluations, with strong driving visual-question-answering results relative to larger general VLMs.
Weights, inference code, and demo data are published on Hugging Face and GitHub under the Apache 2.0 license. The repository layout places the VLM at the root, with planner-sft, planner-rl, and perception heads in adjacent folders for selective download. Developers can run VQA alone, attach perception, or load a planner for bundled demo scenes without proprietary fleets or private logs.
Qwen frames the project as an initial open step toward a unified vision-language foundation for autonomous driving research, inviting reproduction, ablation, and extension rather than presenting a closed production stack.
ANALYSIS The architectural choice to leave the Qwen3.5-4B trunk untouched while bolting on driving-specific heads makes the release modular: downstream teams can swap or retrain individual components without destabilizing the base VLM. Shipping two planner checkpoints, one imitation-learned and one RL-optimized, gives researchers a direct ablation point for comparing training regimes on the same perception backbone. At 4B parameters, the model sits in a size class that can run on consumer-grade GPUs, lowering the hardware barrier for academic groups working on end-to-end autonomous driving.