Alibaba's Qwen Releases Qwen-Drive-1.0-4B, a Unified VLM for Autonomous Driving Perception and Planning

Alibaba's Qwen team has released Qwen-Drive-1.0-4B, a vision-language foundation model for autonomous driving that combines 3D perception, visual question answering, and motion planning within a single architecture. The weights are available on Hugging Face, with code and documentation published at QwenLM/Qwen-Drive-1.0 on GitHub.

What it is

Qwen-Drive-1.0 is built on the pretrained Qwen3.5-4B vision-language model, which the team says remains architecturally unchanged. Two external modules are attached to this shared VLM backbone:

• A BEV perception head that jointly performs 3D object detection, semantic occupancy prediction, and bird's-eye-view (BEV) map segmentation.
• A Planning Expert that conditions on the VLM's representations to generate future ego trajectories using flow matching.

The repository ships as a single directory: the 9.1 GB VLM at the root (which alone handles VQA), plus a 2.1 GB planner-sft module, a 2.1 GB planner-rl module, and a 0.5 GB perception module. Two Planning Experts are released — planner-sft, trained via imitation learning and supporting both direct and reasoning-based planning, and planner-rl, further optimized with reinforcement learning against NAVSIM PDMS, WOD-E2E RFS, and a displacement term.

Benchmark claims

According to Qwen, Qwen-Drive-1.0-RL scores 90.7 on NAVSIM PDMS (91.4 best-of-6) and 8.45/7.91 on WOD-E2E RFS (val/test), positioning it competitively against baselines including AutoVLA, SpanVLA, MindVLA-U1, and Alpamayo-1.5. On driving VQA, the SFT variant reports a LingoQA score of 77.8 (79.4 under the official LingoJudge protocol), an Ego3D RMSE of 7.78, and a WaymoQA-all score of 74.5 — each claimed as the top result among compared models including InternVL3.5-8B, LLaVA-OV2-8B, and Cosmos-Reason2-8B.

On general vision-language benchmarks, Qwen states the SFT model performs on par with base Qwen3.5-4B, reporting 85.5 on MMBench, 75.9 on MMStar, and 72.7 on MMMU, suggesting no significant catastrophic forgetting after driving-specific training. These figures come from Qwen's own technical report and have not been independently verified.

Training approach

Qwen describes a staged training recipe that unifies trajectory annotations and labels across multiple public driving datasets, filters samples for consistency, and blends driving supervision with general-purpose vision-language data. The company states this design allows domain adaptation to driving tasks while mitigating loss of broader instruction-following and visual reasoning capability.

Pricing is not applicable — the model weights are released openly on Hugging Face. Context window and standard per-token pricing figures do not apply to this specialized multimodal driving model in the way they would to a general-purpose LLM API.

What this means

Qwen-Drive-1.0-4B signals continued expansion of general-purpose VLM providers into domain-specific embodied AI applications, following a pattern seen with models like Cosmos-Reason and Alpamayo. By keeping the base VLM architecture untouched and attaching modular perception and planning heads, Qwen is betting that broad multimodal pretraining transfers well to safety-critical driving tasks without costly retraining. The reported benchmark wins are self-reported and use Qwen's own models (Qwen-Plus) as evaluation judges in at least one case, which warrants independent verification before drawing conclusions about real-world driving performance. The release is notable primarily as an open, inspectable reference architecture for combining perception, reasoning, and planning in one model rather than as a production-ready autonomous driving stack.