Qwen-Drive-1.0: A Vision-Language Model for Driving
Qwen released Qwen-Drive-1.0 on September 3, 2026. This model unifies 3D perception and visual question answering using the Qwen3.5-4B base with external…
Release Overview and Architecture
Qwen announced the release of Qwen-Drive-1.0 on September 3, 2026. The source describes this as the first vision-language foundation model for autonomous driving that unifies 3D perception and visual question answering at the pretraining stage. The model further extends these capabilities to motion planning. According to the official changelog, the pretrained VLM architecture remains entirely untouched during this process.
- Base architecture: Natively multimodal Qwen3.5-4B.
- External modules: Two modules are attached to the base model.
- Specific component: A BEV perception head is included among the external modules.
Confirmed Capabilities and Limitations
The source confirms that the model accepts text and image inputs and produces text outputs. It explicitly states that open weights are not confirmed for this release. No pricing information, context window size, or maximum output token limits are provided in the available facts. The documentation does not specify deployment requirements, local installation feasibility, or API availability details beyond the general release announcement.
- Input modalities: Text and image.
- Output modality: Text.
- Open weights status: Not confirmed.
- Pricing: Not specified.
Evaluation Procedure for Engineers
Engineers can evaluate Qwen-Drive-1.0 by designing specific tests aligned with its stated functions. First, test the model’s ability to perform 3D perception tasks using image inputs. Measure the accuracy of the generated spatial representations against ground truth data. Next, validate the visual question answering capability by posing specific queries about driving scenarios and comparing the textual responses to expected answers. Finally, assess the motion planning extension by providing scenario-based inputs and measuring the consistency of the planned trajectories. Compare these results against baseline metrics to determine performance characteristics within the defined scope.
Primary sources
Frequently asked questions
What is the base architecture of Qwen-Drive-1.0?
The model is built on the natively multimodal Qwen3.5-4B architecture. It attaches two external modules, including a BEV perception head, while keeping the pretrained VLM architecture entirely untouched.
Are open weights available for Qwen-Drive-1.0?
The source indicates that open weights are not confirmed for this release. No additional information regarding weight distribution or access methods is provided in the verified facts.