The paper presents Mage-VL-Cowork, a 4B model that combines visual GUI interaction with code and terminal execution in one policy. The authors report 66.0% on OSWorld-Verified.
Training: A five-stage curriculum across midtraining, SFT and RL, starting from a general multimodally aligned checkpoint.
Data: Mage-CUA-Forge synthesizes 293,806 desktop and web tasks across 6,915 skills, which the authors say cuts recurring per-task cost by 82.6%.
Teacher setup: Rather than distilling a default teacher, they evolve the teacher harness and build specialized CUA harnesses to produce trajectories.
Availability: The model is released on Hugging Face as microsoft/Mage-VL-Cowork.
Sign in to comment.
TL;DR of the paper, generated automatically.
The paper presents Mage-VL-Cowork, a 4B model that combines visual GUI interaction with code and terminal execution in one policy. The authors report 66.0% on OSWorld-Verified.