Open foundation stack for computer-use agents (NeurIPS 2025 Spotlight): the AgentNet dataset, annotation tool, and benchmark plus the OpenCUA-7B/32B/72B model family — the strongest open computer-use models on OSWorld-Verified at release. Post-trained on open VLM backbones (not pretrained from scratch), with chain-of-thought-style reflective reasoning over screenshot-action trajectories.

Joint XLANG Lab (HKU) / Moonshot AI co-development, HKU-led; ships under the xlangai HuggingFace org. Companion to the lab's OSWorld evaluation line.

Model Details

Variants

Name Parameters Notes
OpenCUA-7B post-trained on an open 7B VLM backbone
OpenCUA-32B post-trained on an open 32B VLM backbone
OpenCUA-72B post-trained on an open 72B VLM backbone

Paper

Venue NeurIPS 2025
agenticmultimodalopen-weight

Related