MintAct is a family of vision-language models (2B/4B/8B) that unifies UI grounding, multi-step navigation across mobile, desktop, and web, and visual tool use within a single trained model, matching specialist per-domain models on all these capabilities. The authors built a scalable RL infrastructure hosting hundreds of concurrent per-domain environment instances and an asynchronous training framework that keeps explicit control over the cross-domain training distribution while remaining stable under noisy feedback and off-policy drift. MintAct achieves state-of-the-art results (48.9 on OSWorld-Verified) among comparably sized models across a wide range of benchmarks.
