Alibaba’s Qwen Releases Driving Model Whose Explanations Can Diverge From Maneuvers

Alibaba’s Qwen research team has released Qwen-Drive 1.0, an open model designed to understand driving scenes, answer traffic questions and plan a vehicle’s next movements. It gives researchers one system to inspect across those tasks, but its explanations can diverge from the maneuvers it plans.

Qwen is releasing the model free to the research community through Hugging Face, ModelScope and GitHub. It builds on Qwen3.5-4B, an AI model that works with images and text, then adds components for a top-down 3D road view and short-horizon route planning.

One model, three driving jobs

The shared model handles the language and image work. One added component identifies objects in three dimensions, marks usable and blocked areas, and maps road features from a bird’s-eye view. Another uses the model’s outputs to generate the vehicle’s future path.

Qwen’s researchers found that attaching those components alone did not solve the spatial problem. Training only the added pieces left spatial accuracy low; substantial improvement required training the core image-and-text model on spatial tasks too. The training pipeline used 24 public traffic-scene datasets before a reinforcement-learning stage intended to refine driving behavior.

Better simulator behavior came with restraint

The reported improvement also changed how the model drove: it was more cautious and covered less distance in the simulator. Qwen tested planning from simple predictions through a simulator where mistakes can compound, so the result is evidence about that setting rather than real-world driving performance.

The reason may not be the action

Qwen-Drive can explain decisions such as braking or turning. But its explanation can name an incorrect or incomplete cause, and the planned maneuver does not always match the reasoning offered beforehand. A readable rationale, then, is not necessarily a faithful account of the plan.

Its perception also dropped substantially on footage from vehicles with different camera configurations, where suitable training data was limited. Researchers can now test whether the model’s scene understanding and generated reasons hold up when the camera setup changes.