Qwen, the Alibaba-affiliated lab, published Qwen-Drive-1.0 on GitHub, a research model that adds driving perception and trajectory planning onto its existing Qwen3.5-4B vision-language model. The repository, built with Huazhong University of Science and Technology, ships weights on Hugging Face and ModelScope, code on GitHub, and a technical report on arXiv, all under an Apache 2.0 license.
The design keeps the original Qwen3.5-4B decoder untouched, so the base model can still field general visual questions, and attaches two new pieces to it: a perception head that predicts 3D object boxes, occupancy grids and bird’s-eye-view maps, and a planning module that outputs a future path for the vehicle. Qwen trained the system in stages, folding scene perception, question answering and route planning into one process, and built a data pipeline that maps different perception label formats into a shared schema, cleans up driving-related answers for consistency, and converts trajectories pulled from several public driving datasets into one waypoint format.
Two versions of the planner exist. One is trained by imitation, labeled SFT. The other is the same expert further tuned with reinforcement learning against the benchmark objectives. On the self-reported numbers, the RL planner scores 90.7 on the NAVSIM navtest PDMS metric versus 88.2 for the SFT version, while the SFT model actually edges out RL on the NVIDIA PhysicalAI open-loop test, posting a 0.34 meter average displacement error against 0.38 meters for RL.
Qwen also ran driving question-answering comparisons against seven other vision-language systems, including InternVL3.5, LLaVA-OV2 and Cosmos-Reason2, and reports its own SFT model ahead on most of the listed scores. One detail stands out: on the LingoQA benchmark, Qwen says it scored the model with its own judge model, Qwen-Plus, instead of the standard LingoJudge, because the team found the official judge lenient and inconsistent. The repository does separately disclose the score under the standard protocol, 79.4, so the comparison is at least available, just not the one Qwen leads with.
None of this has been checked by anyone outside Qwen. The benchmarks, the judge substitution and the comparison protocol are all set by the lab, run on its own test harness, and posted without independent verification. That is ordinary for a research repository, but it carries more weight here than in most model releases: this is software meant to reason about driving decisions, and a self-graded table of benchmark scores is not the same kind of evidence as a road test. The GitHub page itself frames the release as “an initial step” toward the stated goal, not a finished driving system, and nothing in the repository suggests the model is meant to control a vehicle today.
The detail worth building around is architectural rather than a leaderboard number. Driving-specific models have historically traded away general competence to get good at the driving task, ending up narrow and hard to reuse for anything else. Qwen’s own tables show the driving-tuned model landing close to its unmodified base on general benchmarks such as MMBench (85.5 versus 87.1) and MMMU (72.7 versus 73.4), which is the more interesting claim than any single planning score. A vision-language model that keeps most of its general ability while gaining a dedicated planning head is a different kind of building block than one that only knows how to drive.
The hardware bar is modest for a research setup. Qwen recommends a GPU with at least 24GB of memory to run the released weights, which split into a 9.1GB base model plus separate 2.1GB planner checkpoints and a 0.5GB perception head that attach to it, alongside demo scenes for a quick first run. Teams evaluating vision-language stacks for driving-adjacent work, such as scene description, spatial reasoning or simulation tooling, should treat Qwen-Drive-1.0 as an architecture worth benchmarking independently, not a component to point at real driving decisions until outside evaluation catches up with the paper.
Source: the QwenLM Qwen-Drive-1.0 repository, published by Qwen on GitHub with its accompanying technical report and Hugging Face model card.
