Models

← All models

OpenCUA-7B

XLANG Lab · 2025-08-13 · 8.3B parameters

OpenCUA-7B is an open-weight, 7B-class computer-use vision-language model fine-tuned from Qwen2.5-VL-7B-Instruct and released with the OpenCUA/AgentNet project as a NeurIPS 2025 Spotlight. Its released checkpoint was jointly fine-tuned end-to-end for an additional 200 billion tokens on a mixture of 20% planning, 20% grounding, and 60% general data; the exact cumulative model exposure is not stated because the 200B figure covers this additional stage rather than the pretrained base. The default OpenCUAAgent setup consumes text instructions and GUI screenshots and emits reflective L2 chain-of-thought plus executable actions across Ubuntu, Windows, and macOS.

Benchmark scores