Microsoft releases Mage to combine desktop actions and code
Microsoft has released Mage-VL-Cowork, a computer-use research model that chooses between mouse and keyboard actions and terminal code during a task. Its weights, evaluation code and technical report are available. A single learned policy selects the channel, while the training process brings interface skills and coding together before reinforcement learning.
Artificial Intelligence··Evening
One model chooses between clicking and code
Microsoft released the weights of Mage-VL-Cowork on 30 October. The computer-use research model looks at screenshots and chooses between mouse and keyboard actions and code executed through a terminal. A single learned policy selects the channel at each turn, including switching during the same task. Microsoft’s project page links to the weights, evaluation code and technical report; i-SCOOP independently reported the release and its desktop-and-terminal approach.[1], [2]
Five training stages bring the channels together
The training sequence begins with interface-oriented midtraining and supervised graphical actions. Coding and terminal skills follow, before hybrid demonstrations combine the two channels. Online reinforcement learning then uses verifiable task rewards. The team generates instructions, environment setup and grading machinery together. The experiments run on a Linux desktop, with demonstrations covering spreadsheets, documents, presentations and image editing. The terminal channel works directly with structured files while graphical actions handle visible application state.[1]
Research tools accompany the release
Microsoft also describes WebVoyager-Static, a controlled evaluation using local replicas of websites. Its checks inspect backend state and page visits alongside visible interface evidence. Mage-VL-Cowork is presented as a research checkpoint and training approach, with the developer’s evaluations tied to their stated environments. The published weights allow researchers to examine the model in their own setups. The project’s documentation describes both channel selection and the task-generation process used to train it.[1]
