Autonomous driving has spent two years being pulled forward by end-to-end pipelines, but one old problem still sits in the middle: general-purpose vision-language models cannot read the 3D world, while classical driving stacks have no language or reasoning. On August 31, the Qwen team posted Qwen-Drive-1.0 to arXiv — "An Initial Step towards a Vision-Language Foundation Model for Autonomous Driving" — with 16 authors. Within two days it reached #2 paper of the day on Hugging Face with 335 upvotes: Qwen's first official move to drive its VLM line into the driving domain.

Untouched backbone, two bolt-on modules

The most notable design choice: the pretrained VLM architecture is kept entirely intact. Two external modules are attached instead.

One is a bird's-eye-view (BEV) perception head that jointly performs 3D object detection, semantic occupancy prediction, and BEV map segmentation. The paper frames it as a probe — measuring how much 3D information the shared representations actually contain, and providing an explicit, inspectable interface to 3D scene structure.

The other is a Planning Expert that conditions on shared VLM representations to generate future ego trajectories. First author Xin Zhou added two details in the HF comment section that the abstract omits: the base is the natively multimodal Qwen3.5-4B, and trajectories are generated through flow matching.

Staged training: learn to drive, don't forget to chat

Domain adaptation usually fears catastrophic specialization. Qwen-Drive-1.0 uses a staged training recipe that mixes driving supervision with general-purpose vision-language data, acquiring driving-specific competence while preserving broad visual understanding and instruction-following.

Evaluation spans open-loop, pseudo-closed-loop, and closed-loop settings across 3D perception, driving VQA, and motion planning. The paper's wording is careful: strong 3D perception and driving scene understanding, general capability "largely preserved," and "highly competitive" motion planning.

No weights yet, code "will be" on GitHub

Timeline note: the arXiv page says code will be available at github.com/QwenLM/Qwen-Drive-1.0, and the HF page shows zero linked models — for now you can read the paper but cannot download weights. The first author calls it "the first vision-language foundation model for autonomous driving that unifies 3D perception and visual question answering at the pretraining stage" — an author claim, to be tested by independent evaluation.

So what

The real signal is the paradigm, not the scores: instead of training driving models from scratch, take a general VLM as the base and bolt on domain experts. That is exactly the "foundation model + domain adaptation" direction the industry is converging on. The only question left: will driving become the next battleground for VLM foundation vendors?

Reference: https://arxiv.org/abs/2609.00111