Microsoft just gave away an AI that runs your computer by itself

Microsoft has open-sourced Mage-VL-Cowork, a small 4-billion-parameter model that can operate a computer on its own. The weights went up on Hugging Face yesterday afternoon under the Apache 2.0 license, and today developers are passing it around, especially in Chinese tech circles.

The trick is that the model decides at every step how to act. If a job is easier with the mouse, it looks at a screenshot and clicks through the interface. If code is faster, it opens a terminal and writes a Python script. There is no separate orchestrator agent on top, it all happens inside one model.

The demo is almost cruel. The task in LibreOffice Calc: merge two banner ranges, label them, build a chart and save a new file. Mage-VL-Cowork finished in 39 seconds and 6 steps by editing the workbook with a script. Qwen3.6-27B, about six times bigger and limited to clicking, took 9 minutes 56 seconds and 39 steps. In GIMP the small model edited an image in 1 minute 58 seconds and checked the saved file from the terminal, while its rival gave up after 7 minutes with a file that failed the check.

On OSWorld-Verified, a benchmark where agents do real tasks in desktop apps, the model scores 66% with a 50-step budget. In Microsoft's table humans get 72.4%, and the top GPT-5.5 and Claude Opus 4.7 sit around 78%. So it doesn't catch the flagships, but it beats models several times its size, including Muse Glimmer-30B and Qwen3.6-27B. All the numbers come from Microsoft itself for now, nobody has checked them independently.

Fun detail: the language part of the model is built on China's Qwen3-4B, with Microsoft's own image encoder bolted on. The weights are about 9.5 GB, the suggested way to run it is SGLang, and there is a ready Docker image. This morning we wrote about how Vercel handed your desktop to AI agents, and now the agent itself fits on a home machine.

A computer-use agent no longer needs a giant cloud brain.