Found first: a primary source the press has not covered yet.
The Qwen team submitted Qwen-Drive-1.0 to arXiv on August 31, 2026, describing a vision-language model that performs 3D object detection, semantic occupancy prediction, BEV map segmentation, visual question answering, and ego-trajectory planning from a single set of shared representations. The paper's full title is "Qwen-Drive-1.0: An Initial Step towards a Vision-Language Foundation Model for Autonomous Driving," and it lists 16 authors led by Xin Zhou and Zongchuang Zhao.
What the source says
The architecture has three named components. A BEV Perception Head handles 3D object detection, semantic occupancy prediction, and BEV map segmentation simultaneously. A Planning Expert module generates future ego trajectories conditioned on the shared VLM representations rather than on a separate perception stack. A staged training recipe combines driving-specific supervision with general vision-language data, which the authors say preserves general VL capabilities while adding driving competence. The model is evaluated in open-loop, pseudo-closed-loop, and closed-loop settings; the abstract does not report specific metric values. Code is promised in a forthcoming GitHub release under the Qwen-Drive-1.0 repository.
Why it matters
Most deployed autonomous driving stacks treat perception, language understanding, and planning as separate modules with separate training pipelines. Qwen-Drive-1.0 proposes a single model where the same representations feed all three tasks, which is a different architectural bet. The Qwen lineage has produced a series of consequential open-weight releases across vision and language, and this paper extends that lineage directly into the driving domain. Whether the shared-representation approach holds up in closed-loop evaluation against specialized pipelines is the open question the paper raises but, by its own admission, does not fully resolve.