Microsoft Releases 4B Mage-VL-Cowork Computer Use Model

Microsoft detailed Mage-VL-Cowork in a model card, describing a 4 billion parameter vision language model that can operate graphical interfaces and execute code through a persistent terminal. The model selects between computer use actions and Bash commands without an external router.

Mage-VL-Cowork scored 66.0% on OSWorld Verified under a 50 step limit, based on the average of two runs across 361 tasks. Microsoft reports scores of 87.5% on operating system tasks and 62.7% on workflows spanning multiple applications.

Microsoft trained the model through a five stage curriculum covering visual grounding, graphical interface use, coding, terminal work, supervised hybrid training, and online reinforcement learning. The reinforcement learning stage increased accuracy from 55.84% to 66.0%, while reducing average model calls by 13.0% and output tokens by 25.4%.

The model uses a 32,768 token active context window and retains the three most recent screenshots. Its weights are available on Hugging Face under the Apache 2.0 license. The benchmark results are Microsoft's own and have not been independently verified.