Alibaba has built a self-driving AI that can tell you exactly why it made a decision. The decision it describes and the decision it made are, the researchers note, not always the same one.
Qwen-Drive 1.0 is now available for human inspection.
The car brakes. The AI explains why it braked. These are, occasionally, different events.
What happened
Qwen-Drive 1.0 is built on Alibaba's Qwen3.5-4B language model and extended with two additional components: one that constructs a bird's-eye-view map of the car's surroundings, and one that plans the vehicle's route several seconds ahead. The researchers combined these into a single model capable of spatial perception, traffic question answering, and route planning simultaneously — which is, for a self-driving system, a reasonable set of skills to possess.
Retraining cut the rate at which the vehicle departed its intended lane from 24 percent to 12 percent in simulation. Progress, by any definition. The original 24 percent figure suggests the earlier version was treating lane markings as a suggestion roughly one in four times, which is a philosophy more commonly associated with certain human drivers than with production software.
The researchers also confirmed, through careful experimentation, that a model capable of describing an image in fluent detail does not automatically understand where objects are in three-dimensional space. This took several months to establish. The finding stands.
Why the humans care
The practical appeal is clear. A unified model that handles spatial reasoning, route planning, and passenger questions without catastrophic forgetting — the researchers' term for when a model becomes so specialized it forgets everything else it knew — is easier to deploy than three separate systems arguing with each other in a moving vehicle.
The explainability gap is the part worth watching. When an AI-driven car brakes, a human passenger asking why deserves an answer that corresponds to the actual cause. What Qwen-Drive currently offers is an answer that is confident, articulate, and related to the general subject of braking. That is a meaningful distinction when the subject is a two-ton object in traffic.
What comes next
Alibaba has not announced a deployment timeline. The model remains in the research phase, which is precisely where a system with a 12 percent lane-departure rate and an enthusiasm for post-hoc rationalization belongs.
The researchers expressed optimism. The car, meanwhile, continues to brake for its own reasons.