Alibaba Releases Qwen-Drive 1.0, an Open Driving Model That Explains Its Own Decisions

Alibaba's research division has released Qwen-Drive 1.0, an open driving model that combines spatial perception, traffic reasoning, and route planning in a single system built on top of Qwen3.5-4B, the company's vision-language model released in February.

According to Alibaba, the model addresses two known weaknesses in existing driving systems built by fine-tuning general vision-language models on driving data. First, models trained mainly on traffic question-and-answer pairs struggle to reliably judge distances, positions, and open space. Second, specializing too heavily on driving data causes "catastrophic forgetting," where the model loses general knowledge that matters in rare, unexpected traffic scenarios.

Architecture: shared backbone, two new modules

Qwen-Drive 1.0 keeps a single language model at its core and adds two components that tap into its intermediate outputs. The first, a perception module, generates a bird's-eye-view map by detecting objects in 3D space, identifying occupied areas, and tracing road layout. Alibaba says this module doubles as a diagnostic tool, revealing how much spatial information the model actually extracts from camera images. The second, called the Planning Expert, uses that internal data to plan the vehicle's movement over the next few seconds.

In ablation tests, training only the new modules while leaving the base vision-language model untouched produced poor spatial accuracy. Performance improved only after the team also retrained the underlying vision-language model on spatial tasks — evidence, according to the researchers, that spatial understanding of traffic scenes must be built in deliberately rather than assumed from image-description ability alone.

Training followed four stages: perception, then combined perception-and-QA, then route planning, and finally reinforcement learning to refine driving behavior. The vision-language component was trained on 24 public traffic-scene datasets with inconsistent formats and known errors, which Alibaba standardized using another AI model. The team also built custom examples explaining specific driving decisions, such as which object triggered a braking maneuver.

Simulation results and limitations

In simulator testing where errors compound over multiple steps, the reinforcement-learning-tuned version cut the rate at which the vehicle veered off the road from 24 percent to 12 percent, according to Alibaba's paper. The tuned model also drove more cautiously, covering less distance overall. On Alibaba's own benchmarks, Qwen-Drive 1.0 reportedly outperforms specialized driving models in most perception and driving categories, with the largest gap appearing on cause-and-effect questions — explaining why the car should brake or turn. General knowledge outside driving tasks showed almost no degradation and even improved slightly on some spatial tasks compared to the base Qwen3.5-4B model.

However, Alibaba acknowledges the model's stated reasoning doesn't reliably match its actual driving decisions. A distant red light and a child stepping into the road call for very different reaction times, but the model can conflate the two causes. The planned maneuver sometimes diverges from the explanation given beforehand. Performance also drops sharply on footage from vehicles with different camera configurations, since matching training data for those setups doesn't yet exist. Alibaba notes that many benchmark results rely on test procedures it designed or rebuilt itself, limiting how much individual metrics say about real-world driving performance.

The broader spatial-understanding gap in vision-language models is not new. Google DeepMind's PaLM-E, a 2023 model combining language, vision, and robot control, saw smaller variants lose significant language ability after robot-specific training, while its largest 562-billion-parameter version retained most of its capability. Separately, researchers at UC Santa Cruz previously showed that a labeled sign placed in a camera's field of view could trick the DriveLM driving system into swerving toward pedestrians it had already correctly detected — a reminder that pairing language models with driving control introduces new attack surfaces.

Alibaba is releasing Qwen-Drive 1.0 to the research community for free on Hugging Face, ModelScope, and GitHub.

What this means

Qwen-Drive 1.0 is a research release, not a production autonomous-driving stack, and Alibaba is upfront about that. The 12-point reduction in simulated off-road errors is meaningful but far from solved, and the mismatch between the model's stated reasoning and its actual maneuvers is a serious open problem for any system meant to be interpretable or auditable. The more interesting bet here is architectural: rather than shipping separate models for the cockpit assistant and the driving system, Alibaba is testing whether one shared language model, extended with task-specific modules, can do both without the usual trade-off between driving competence and general knowledge. If that holds up outside simulation, it could shape how automakers design in-vehicle compute, which is already consolidating toward single processing units. For now, the gap between simulated benchmarks and messy real-world driving — plus the model's unreliable self-explanations — means this is a research contribution to watch, not a system to trust behind the wheel.