ScaleCUA-qwen3.5-osworld

ScaleCUA-qwen3.5-osworld

A Qwen3.5-9B computer-use agent from ScaleCUA, trained for OSWorld. Stage: online-RL (RLVR).

• 📄 Paper: SCALECUA: Scaling Computer Use Agents with Verifiable Task Synthesis and Efficient Online RL
• 💻 Code: https://github.com/THUDM/SCALE-CUA
• 🤗 Collection: https://huggingface.co/collections/extreme1228/scalecua-6a4cad5e3c80a230622eaff3 · 🤖 ModelScope

What is this

This is the online-RL checkpoint (the released agent). Its SFT start point is extreme1228/ScaleCUA-qwen3.5-osworld-sft. It reaches 68.7% on OSWorld (open-source SoTA).

ScaleCUA scales online RL for computer-use agents by pairing VeriGen (verifiable GUI-task synthesis) with Frontier Sampling and Visual Context Segmentation. See the repo for details.

Usage

Serve with an OpenAI-compatible server (e.g. vLLM) and drive it with the evaluation runtime in osworld_eval:

pip install -U vllm
python -m vllm.entrypoints.openai.api_server --model extreme1228/ScaleCUA-qwen3.5-osworld \
  --served-model-name scalecua --trust-remote-code --port 8000

Then follow REPRODUCE.md to run OSWorld evaluation.

Results

68.7% on OSWorld (open-source SoTA) (see the paper for the full table).

License

Weights are released for research; the model inherits the license of its base model (Qwen3.5-9B). Repository code is under the licenses in the GitHub repo.

Citation

@misc{lv2026scalecua,
  title  = {ScaleCUA: Scaling Computer Use Agents with Verifiable Task Synthesis and Efficient Online RL},
  author = {Bowen Lv and Xiao Liu and Yanyu Ren and Hanyu Lai and Bohao Jing and Hanchen Zhang and Yanxiao Zhao and Shuntian Yao and Jie Tang and Yuxiao Dong},
  year   = {2026},
  url    = {https://github.com/THUDM/SCALE-CUA}
}