CUA-Master-27B scores 80.2 on OSWorld Verified after training on video-mined workflows
Computer-use agents operate apps by clicking and typing. CUA-Master mines real workflows from instructional videos as seed tasks, turns them into reusable pieces to build new verifiable tasks, and retrains on the model's weak spots in repeated rounds. The 27B model scores 80.2 on OSWorld Verified, 36.9 on OSWorld 2.0, 43.9 on WeaveBench and 25.0 on Agents' Last Exam.