# Mage-VL-Cowork trains a 4B model for computer use

> The authors report 66.0% on OSWorld-Verified for a model trained to combine screen interaction with code and terminal use.

Oossa · 2026-11-02 · https://oossa.com/en/mage-vl-cowork-trains-a-4b-model-for-computer-use

Researchers have introduced Mage-VL-Cowork-4B, a computer-use model designed to combine visual interaction with apps and websites with code and terminal commands. Their training system generated 293,806 desktop and web tasks across 6,915 skills; the authors say it cut recurring task-generation costs by 82.6%.

The team reports a 66.0% score on OSWorld-Verified and calls it state of the art among models of its size. That result is reported by the authors; the abstract does not provide independent validation or details about access to the model.

## The facts

- Model size: 4 billion parameters; reported OSWorld-Verified score: 66.0%.
- Training data: 293,806 tasks spanning 6,915 skills.

## Why it matters

For developers exploring computer-use agents, the reported score suggests a small model can handle a range of desktop and web tasks, but the abstract does not establish how it performs in independent or real-world use.

## Sources & references

1. [Mage-VL-Cowork: A Full Training Recipe to Learn Hybrid Computer Use From Evolved Harnesses](https://arxiv.org/abs/2610.40747) – arXiv, 2026-11-02

Last updated: 2026-11-02