Updates·October 31, 2026, 16:50

Microsoft unveils small AI model that controls the computer on its own

AI-generated from the sources listed, with automated source checks.

Microsoft has released Mage-VL-Cowork, a relatively small AI model that can use the screen like a human and also run code. According to Microsoft, it outperforms other models of the same size on a computer-task benchmark.

AI-generated image

Microsoft has unveiled Mage-VL-Cowork, an AI model that can control a computer almost like a human. It can see and use programs through the screen while also writing and running code and commands in a terminal. All of this happens in a single model with 4 billion parameters. Parameters are the settings a model learns during training, and the number is a rough measure of the model's size.

Results according to Microsoft

On the OSWorld-Verified benchmark, the model scores 66.0 percent, according to Microsoft, when it is allowed at most 50 steps per task. Microsoft calls it the highest result among models with 4 billion parameters. The sub-scores are 87.5 percent on operating system tasks and 62.7 percent on tasks that span multiple programs.

All figures are Microsoft's own and have not been independently verified.

How the model is built

Mage-VL-Cowork is built on Microsoft's foundation model Mage-VL. It uses the Mage-ViT vision component and the Qwen3-4B language model. The model has a context window of 32,768 tokens, meaning how much text it can keep track of at once. Its memory weights the most recent screenshots most heavily.

Training takes place in five stages, including fine-tuning and reinforcement learning, in which the model learns from attempts and outcomes. The process moves from learning graphical user interfaces, to mastering the terminal, to combining the two. Training data was created with the Mage-CUA-Forge tool, which, according to Microsoft, produced 293,806 tasks across 6,915 skills, as well as 318,837 mixed trajectories.

Microsoft states that training with online hybrid reinforcement learning yields 10.16 percentage points more than fine-tuning alone. At the same time, the number of model calls falls by 13.0 percent and the number of output tokens by 25.4 percent.

For research only and with caution

The model is released under the Apache 2.0 license and is available on Hugging Face. According to the README file, it is intended for research use and is not meant for deployment in products or services.

Microsoft writes that the model was designed and tested in isolated sandbox environments that can be discarded. It requires strict security, including container isolation and human approval of actions that cannot be undone.

Mage-VL-Cowork is not the same as Microsoft 365 Copilot Cowork, which is a different product. The exact release date is not stated in the sources.

Sources

More on this topic

Get the week's AI news in your inbox

Choose your level, topics and length. One email a week, unsubscribe at any time.

Subscribe to Promptly Newsletter