Models

← All models

OpenCUA-32B

XLANG Lab · 2025-08-13 · 33.5B parameters

OpenCUA-32B is an MIT-licensed open-weight end-to-end computer-use model from the OpenCUA project, derived from Qwen2.5-VL-32B-Instruct through two training stages reported as 35 billion and 60 billion tokens on mixtures of grounding, general text and vision, and GUI-trajectory data. Its agent setup uses three screenshots of visual history and L2 reflective chain-of-thought at inference; the release paper reported 34.8% average success on OSWorld-Verified at a 100-step budget, then state of the art among open-source models.

Benchmark scores