What changed
Qwen released Qwen-Drive-1.0, a four-billion-parameter vision-language foundation model that combines driving visual question answering, 3D perception and motion planning. Apache-2.0 weights, code, demo data, documentation and a technical report are public.
Why it matters
The release gives autonomous-driving researchers an inspectable model and training/evaluation artifacts across perception, language and planning. Open weights improve reproducibility but do not establish road safety, closed-loop robustness or production readiness.
What to do
Use it only in simulation or controlled research. Pin the model and code revision, reproduce the published open-loop and closed-loop evaluations, test rare and out-of-distribution scenes, and require a separate safety case before any vehicle-control experiment.
Caveat
The reported performance is author-run, one listed VQA metric uses Qwen-Plus rather than the benchmark's official judge, and the project describes itself as an initial research step. The public artifacts are not approval for real-world autonomous operation.