ToolCUA-8B is an end-to-end computer-use agent for orchestrating GUI actions and structured tool calls. It learns when to continue through GUI interaction, when to invoke tools, and when to switch back, enabling shorter and more reliable desktop task trajectories.
We recommend vLLM for deployment. Use vllm>=0.12.0 and enable --trust-remote-code.
1MAX_IMAGE=${MAX_IMAGE:-5}
2IMAGE_LIMIT_ARGS='{"image": '"$MAX_IMAGE"'}'
3PIXEL_ARGS='{"size": {"longest_edge": 3072000, "shortest_edge": 65536}}'
4
5vllm serve X-PLUG/ToolCUA-8B \
6 --trust-remote-code \
7 --max-model-len 32768 \
8 --mm-processor-kwargs "$PIXEL_ARGS" \
9 --limit-mm-per-prompt "$IMAGE_LIMIT_ARGS" \
10 --tensor-parallel-size 1 \
11 --allowed-local-media-path '/' \
12 --port 4243 \
13 --gpu-memory-utilization 0.85 \
14 --mm-processor-cache-gb 0 \
15 --no-enable-prefix-caching \
16 --enforce-eager \
17 --max-logprobs 50
1@article{hu2026toolcua,
2 title={ToolCUA: Towards Optimal GUI-Tool Path Orchestration for Computer Use Agents},
3 author={Hu, Xuhao and Zhang, Xi and Xu, Haiyang and Qiao, Kyle and Yang, Jingyi and Huang, Xuanjing and Shao, Jing and Yan, Ming and Ye, Jieping},
4 journal={arXiv preprint arXiv:2605.12481},
5 year={2026}
6}
