← All models

Qwen-Drive-1.0

Published Publisher Alibaba

Overview

Qwen-Drive-1.0 is a vision-language foundation model for autonomous-driving research. It builds on the multimodal Qwen3.5-4B backbone and combines 3D perception, visual question answering and motion planning without changing the pretrained vision-language architecture. The official announcement is dated September 3, 2026. (qwen.ai)

Capabilities and uses

  • Processes single-view and multi-view driving images, image sequences and general images for scene understanding and visual question answering.
  • An external bird’s-eye-view perception head performs 3D object detection, semantic occupancy prediction and map segmentation.
  • A Planning Expert generates five-second future ego-vehicle trajectories through flow matching. The release includes supervised-fine-tuned and reward-optimized planning experts.
  • Intended uses include driving-scene analysis, perception and planning research, and adaptation to driving scenarios. The authors report evaluations in open-loop, pseudo-closed-loop and closed-loop settings. (qwen.ai)

Access and limitations

The official 4B model weights and code are available under the Apache 2.0 license. Planning and perception require their respective external modules; the repository recommends a GPU with at least 24 GB of memory. The authors note that causal reasoning can be unstable, generated trajectories do not always match the textual rationale, and transfer to unseen camera configurations requires further quantitative evaluation. It should not be treated as a validated, stand-alone autonomous-driving system. (github.com)

Pricing

Official input and output API prices: unverified. On November 16, 2026, the model was not found in the checked Alibaba Cloud mainland CNY or international USD model-pricing pages. This does not establish that no service exists; it means no model-specific official price could be verified from those pages. Self-hosting may incur infrastructure costs. (help.aliyun.com)

Official sources