Alibaba’s Qwen team and Huazhong University of Science and Technology have released an open model that attempts to combine three tasks usually handled separately in autonomous-driving research: seeing the road, answering questions about it, and planning a vehicle’s next movement. The project, Qwen-Drive-1.0-4B, became available on September 7 through an official Qwen GitHub repository. It is an early research-oriented foundation model, not a declaration that an autonomous-driving system is ready for deployment on public roads.
The project is noteworthy because its makers have released code, weights, demo data, evaluation material, and a technical report under an Apache 2.0 license. That makes the model accessible to developers who want to inspect or build on its approach. TechNode reported that the release brings Alibaba’s Qwen model family into an area where Chinese companies are competing to connect multimodal models with driving software.
Qwen-Drive-1.0-4B is built on the Qwen3.5 vision-language architecture. The name identifies a four-billion-parameter model, but parameter count alone does not explain the project. Its core proposition is a shared visual-language foundation model with dedicated modules for 3D perception and driving decisions.
One Architecture for Perception, Questions, and Planning
The official repository says Qwen-Drive integrates 3D perception, driving visual question answering, and motion planning in one framework. A bird’s-eye-view perception head is used for 3D object detection, semantic occupancy prediction, and map segmentation. A separate planning expert uses the shared representation to produce future trajectories for the vehicle.
That structure addresses a common engineering problem. Driving systems often split visual understanding, map interpretation, and motion planning into specialized components. Qwen-Drive does not eliminate specialization, since it still uses separate modules. Instead, it anchors the modules in a shared vision-language model so that visual interpretation and driving reasoning can be trained with related data.
The project describes two planning variants. One is trained by supervised fine-tuning, while the other is optimized with rewards after the initial imitation-training stage. The repository says both use the same vision-language model, while the planning expert generates trajectories. In the provided example, the model handles six trajectories with 50 steps, representing five seconds of motion at 10 Hz. Those details explain the research design; they are not proof that the system can safely operate a vehicle in unrestricted traffic.
The release also preserves a general visual-question-answering function in the language-model decoder. That makes the project different from a narrow control stack. It can be used to ask what is happening in a driving scene as well as to produce an intended path. The combination is important for research into systems that need both an interpretable description of the scene and a proposed action.
Open Weights Make Evaluation Easier, Not Automatic
Qwen-Drive’s Apache 2.0 license is central to the release. Developers can obtain the model through GitHub, Hugging Face, and ModelScope, while the repository provides a setup path and demo scenes. The package is substantial: the base vision-language model is listed at 9.1 GB, with separate planning and perception components alongside it. The repository recommends a GPU with at least 24 GB of memory.
Open availability can help researchers test different data pipelines, compare planning approaches, and inspect the connection between visual understanding and trajectory generation. It does not remove the need for independent validation. Autonomous-driving benchmarks have different assumptions, datasets, and evaluation protocols. A strong score under one setup does not establish performance under every weather condition, city layout, hardware configuration, or safety regime.
The repository reports a 90.7 PDMS score for the reward-optimized planner on the NAVSIM v1.1 navtest, as well as multiple scores for driving question answering and spatial understanding. These are the project’s own evaluation tables. They should be read as claims by the research team, not as independently replicated results. The repository itself notes that one LingoQA metric was judged with Qwen-Plus rather than the official LingoJudge because the team considered the official judging protocol inconsistent.
That disclosure is useful because it identifies a methodological choice that users should consider. It also illustrates why open code and reports matter. A developer can examine the evaluation setup instead of relying only on a launch headline.
Alibaba Extends Qwen From Screens to Driving Scenes
The new release fits a wider Qwen strategy of adapting a large model family to task-specific environments. EastFrontier previously covered Qwen UI Agent’s screen-control capabilities, which focused on a model acting across software interfaces. Qwen-Drive shifts the attention from applications on a screen to reasoning about vehicle-camera data, maps, and potential motion.
It also follows a period in which Alibaba has updated the model family for coding and agents. The September upgrade to Qwen3.8-Max emphasized developer use cases. Qwen-Drive instead uses a Qwen3.5 base and targets a specialized research problem. The contrast shows that Qwen is becoming a brand for a collection of models and components, rather than a single general-purpose chatbot.
For China’s autonomous-driving industry, the appeal of this kind of release is not that one four-billion-parameter model replaces an entire driving stack. It is that an open project can give universities, suppliers, and developers a common starting point for experiments in 3D perception, visual reasoning, and planning. It may also create a reference implementation for comparing Chinese and international research systems.
The risks are equally clear. A model can reason convincingly about a driving scene and still make unsafe choices. Benchmark outcomes can fail to capture rare events, sensor errors, or the legal and operational requirements of real vehicle deployment. The official repository does not claim that Qwen-Drive is a production-certified system, and the article should not imply otherwise.
Qwen-Drive’s significance lies in the combination of access and scope. It offers a visible, inspectable attempt to join language-guided understanding with driving-specific perception and trajectory planning. Its next contribution will depend on whether developers reproduce the reported results, identify limits, and adapt the components to more demanding research settings. For now, Alibaba and HUST have given the autonomous-driving research community an open model to test rather than only a proprietary claim to assess.