# Microsoft’s Mage-VL-Cowork is a small open model that decides when to click and when to code

URL: https://ai-pulse.aglaganova.com/blog/microsofts-mage-vl-cowork-is-a-small-open-model-that-decides-when-to-click-and-when-to-code
Published: 2026-11-04
Author: Roman Aglaganov, Full-stack developer
Category: Models
Region: Global
Tags: agents, open-weights

> The 4-billion-parameter agent operates a computer both through the screen and through the terminal, and Microsoft reports 66% on the OSWorld-Verified benchmark. It is released under the Apache 2.0 licence.

![Illustration: a black laptop outline whose screen is split between grey window shapes and grey lines of code, with a small red cursor between them](https://ai-pulse.aglaganova.com/api/media/file/mage-vl-cowork-1.png)

Most AI agents that use computers work like a person staring at a screen: they take a screenshot, decide where to click and repeat. That is slow, because every screenshot costs thousands of tokens, and clumsy for jobs like editing a spreadsheet, which a few lines of code would do faster and more reliably. Microsoft’s Mage-VL-Cowork, released on Hugging Face on 30 October, is trained to choose between the two.

## What it is

Mage-VL-Cowork is a vision-language model with 4 billion parameters, built from Microsoft’s Mage-VL, which combines its own image encoder with Alibaba’s Qwen3-4B. At each step it can either act on the screen, with 11 basic actions such as clicking and typing, or run commands in a terminal and Python. Its model card says it switches by itself, using the screen for visual layouts and the terminal for structured edits to documents, spreadsheets and presentations, with no separate orchestrating program.

## How it was trained

Microsoft first had a large teacher model, which the card names as GPT-5.4, do tasks in an improved setup, then trained the small model on the results through five stages, finishing with reinforcement learning. A synthesis system generated 293,806 practice tasks across 6,915 skills.

## The claims

Microsoft reports 66.0% on OSWorld-Verified, a benchmark of real desktop tasks, which it says is the best result for a model of this size, with 87.5% on operating-system tasks. That figure is Microsoft’s own, averaged over two runs on 361 of the benchmark’s tasks, with those involving Google Drive left out. It says the final training stage also cut the number of model calls per task by 13% and output tokens by a quarter.

## Running it safely

The model is released under the Apache 2.0 licence, and its weights take up about 9.5 GB. Microsoft is explicit that it is an agent that can run arbitrary commands, and says it should only be used inside an isolated container or virtual machine, with no personal keys or passwords, restricted internet access and human confirmation before anything irreversible.

## Sources

- [Hugging Face: microsoft/Mage-VL-Cowork model card](https://huggingface.co/microsoft/Mage-VL-Cowork)
- [Microsoft: Mage-VL-Cowork project page](https://microsoft.github.io/Mage/cowork/)