that only navigate by pixels and mouse clicks are slow, missing shortcuts like terminal commands. Training an agent to choose between mouse and code for each step is challenging.
The researchers built a four-billion-parameter agent, Mage-VL-Cowork, through five fine-tuning stages: first on screenshots, then on coding, and finally on hybrid use. They improved a teacher model by refining its tools before collecting demonstrations, and created a generator that produces thousands of tasks at low cost.
On the OSWorld-Verified benchmark, the agent scored 66%, outperforming models of its size and matching much larger ones. This makes efficient on-device computer assistants more practical.