Qwen-Drive 1.0 puts 3D perception and motion planning in one 4B model

Qwen has published the open-weight Qwen-Drive-1.0-4B model card, marking a substantial move from general-purpose vision-language models toward autonomous-driving systems. The release combines driving-scene question answering, 3D perception, and motion planning around the Qwen3.5-4B vision-language architecture.

The technical report on arXiv describes the same system as an initial vision-language foundation model for autonomous driving. Its scope is broader than visual captioning: the model exposes structured scene information and generates future ego trajectories for driving tasks. The weights and model configuration are available through Hugging Face under the Apache 2.0 license.

Qwen-Drive connects perception, reasoning, and planning

The release keeps the pretrained Qwen3.5-4B VLM as the shared representation. An external bird’s-eye-view (BEV) perception head performs 3D object detection, semantic occupancy prediction, and BEV map segmentation. That explicit head gives developers a visible interface to the model’s spatial representation instead of reducing the entire driving scene to free-form text.

The unchanged VLM can still answer general questions about images and driving scenes. Qwen’s staged training recipe combines driving supervision with general vision-language data, aiming to add driving-specific competence while retaining general visual understanding and instruction-following behavior.

Two planning experts target different operating modes

Qwen also publishes two Planning Expert variants. The planner-sft model supports direct and reasoning-based planning, while planner-rl is further reward-optimized for driving evaluations. Both condition on the shared VLM representation and generate future ego trajectories through flow matching.

This separation matters for practical experimentation. Researchers can inspect perception outputs, compare planning behavior, and evaluate how a common multimodal backbone behaves when paired with different planning heads. It also makes the release easier to study than a monolithic driving model whose intermediate decisions remain opaque.

The release is usable locally, with substantial storage needs

The Hugging Face repository lists the base Qwen-Drive-1.0-4B component at roughly 9.1 GB. The two planner directories add about 2.1 GB each. The model card includes loading paths for Transformers, vLLM, SGLang, and Docker Model Runner, although the published examples still require a suitable local accelerator and runtime setup.

The repository is not a turnkey autonomous vehicle stack. A model card and technical report do not establish road safety, regulatory compliance, or deployment readiness. Closed-loop benchmark results also cannot substitute for validation on the sensors, maps, traffic conditions, and fallback procedures of a real vehicle. Treat the release as a research artifact for reproducible evaluation and system prototyping.

Qwen-Drive’s evidence and open questions

The accompanying arXiv paper reports strong 3D perception and driving-scene understanding, plus competitive results across open-loop, pseudo-closed-loop, and closed-loop motion-planning evaluations. Those claims are based on the authors’ evaluation setup and should be checked against the paper’s datasets, metrics, decoding settings, and released code when it becomes available.

The immediate significance is architectural: one small multimodal backbone now connects visual question answering, inspectable 3D scene structure, and trajectory generation in a public research release. The next milestone is independent reproduction of its planning and safety results, especially outside the benchmark conditions used by the authors.

For workflows that combine models, tools, and verification steps, see LinkLoot’s AI workflow automation guide.

Sources and methodology

The Hugging Face Qwen-Drive-1.0-4B model card is the primary source for the published artifact, license, components, runtime paths, and repository metadata. The independent arXiv technical report corroborates the model’s unified 3D perception, visual question answering, and motion-planning design. Claims about performance remain attributed to the authors and are not presented as independent safety validation.