Apppublic

hugging-apps/qwen-drive-1-vqa-demo

sourceHugging Faceapache-2.0updated 24d agoView on Hugging Face

2likes

App README

Qwen-Drive-1.0-4B

Demo of `Qwen/Qwen-Drive-1.0-4B` — a Qwen3.5-4B vision-language model extended into a foundation model for autonomous driving. Two of the model's three modes are exposed here:

Tab	What runs	Output
Driving VQA	the VLM alone, on an image you upload	text answer
Motion planning	the VLM writes a rationale, then the reward-optimized Planning Expert (`planner-rl`) samples trajectories from its attention cache by flow matching	5 s of `(x, y, heading)` waypoints at 10 Hz, plotted in the ego frame

The BEV perception head is not included — it needs custom CUDA kernels (ms_deform_attn_bf16, voxel_pool) built from source.

Inference follows the authors' reference implementation: the `qwen_drive` package is vendored verbatim into this Space (qwen_drive/, Apache-2.0, see QWEN_DRIVE_LICENSE), and decoding uses the released VQA_DECODE_DEFAULTS and flow-matching sampler settings. The only deviation from scripts/demo.py is attn_implementation="sdpa" instead of flash_attention_2 (matching scripts/run_vqa.py), because FlashAttention needs a source build on this runtime.

Example assets

The four planning scenes and the VQA example images in demo_data/ are the demo scenes bundled in QwenLM/Qwen-Drive-1.0 under the Apache-2.0 license; the frames originate from the Waymo Open Dataset end-to-end driving split. Planning is only offered on these scenes because it requires the full scene record (three camera views at four timestamps, 1.5 s of ego history, and the navigation command); every frame is shown in the gallery before you run it.

Citation

bibtex

@misc{zhou2026qwendrive10initialstepvisionlanguage,
      title={Qwen-Drive-1.0: An Initial Step towards a Vision-Language Foundation Model for Autonomous Driving},
      author={Xin Zhou and Zongchuang Zhao and Zhibo Yang and Mingsheng Li and Humen Zhong and Shuai Bai and Du Chu and Ruizhe Chen and Zhaohai Li and Jun Tang and Qiuyue Wang and Mingkun Yang and Jiazhao Zhang and Dayiheng Liu and Dingkang Liang and Xiang Bai},
      year={2026},
      eprint={2609.00111},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2609.00111},
}