Qwen team publishes Qwen-Drive-1.0, a vision-language foundation model for autonomous driving

Alibaba's Qwen team has released Qwen-Drive-1.0, described as an initial step toward a single vision-language foundation model that can handle 3D perception, visual question answering, and motion planning for autonomous driving in one unified system. The work was published as a paper on arXiv on August 31, 2026, with open weights released under an Apache 2.0 license.

What's new

Qwen-Drive-1.0 is built on top of Qwen3.5-4B, the team's natively multimodal vision-language model, with two additional modules attached to the shared backbone:

• A bird's-eye-view (BEV) perception head that produces explicit, inspectable 3D scene predictions directly from the model's internal representations.
• A Planning Expert that consumes those same representations to generate future ego-vehicle trajectories.

The paper's abstract states: "We present Qwen-Drive-1.0, an initial step towards a vision-language foundation model for autonomous driving. Qwen-Drive-1.0 retains the architecture of the pretrained vision-language model (VLM) and integrates 3D perception, visual question answering, and motion planning within a unified framework."

On 3D perception benchmarks, the model reaches 43.95 mAP and 60.99 map mIoU on nuScenes, and 43.45 mAP and 71.27 map mIoU on OpenScene -- figures the team says are competitive with dedicated vision-based 3D detectors purpose-built for those tasks. On driving-scene understanding, Qwen-Drive-1.0 outperforms the general-purpose Qwen3.5-4B it's built on, while the team reports it retains that base model's general capabilities rather than trading them away for driving-specific skill.

The model and weights are available on Hugging Face and ModelScope.

Context

Autonomous driving has historically relied on stacks of specialized models -- separate systems for perception, prediction, and planning, stitched together with hand-engineered interfaces. The industry has been moving toward end-to-end and VLM-based approaches that fold more of that pipeline into a single learned system, on the theory that a model with broad visual and language understanding can generalize better to rare situations than a narrow perception stack. Qwen-Drive-1.0 extends that trend by starting from an already-capable general-purpose VLM rather than training a driving-specific model from scratch, then attaching lightweight perception and planning heads on top.

It also marks a new application domain for the Qwen family, which has so far been best known for its general-purpose language and multimodal chat models (Qwen3.5, Qwen3.8) and specialized releases like Qwen-Image and Qwen-Image-Edit. Qwen-Drive-1.0 is explicitly framed by its authors as an initial step, language that signals a research release rather than a production-ready driving stack.

Why it matters

For the broader open-weights ecosystem, this is a signal that foundation-model labs are starting to push VLM architectures into safety-critical, domain-specific applications rather than keeping them confined to chat and general reasoning. If a single pretrained VLM can be adapted to perception and planning with lightweight added heads, it lowers the barrier for other teams to build driving-relevant models without training a bespoke architecture from the ground up.

The explicit "initial step" framing is also worth noting: the authors are not claiming a deployable autonomous-driving system, and the benchmark results -- competitive with, not superior to, specialized detectors -- suggest this is best read as evidence that the unified VLM approach is viable, not that it has surpassed dedicated driving-perception architectures yet.