# Microsoft releases Mage-VL-Cowork, a 4B open model that operates a computer through both the screen and the terminal

> OnPrem World · https://onpremworld.com/en/news/2026-10-microsoft-mage-vl-cowork · CC BY 4.0

Published: 2026-10-30 · Models

Microsoft published Mage-VL-Cowork on Hugging Face on October 30 under the Apache-2.0 license. The 4B vision-language model combines screenshot-based mouse and keyboard control with a bash channel in a single checkpoint, choosing between them turn by turn with no external orchestrator. Its weights take 9.48 GB in bfloat16 and it uses a 32,768-token context window. Microsoft reports 66.0% on OSWorld-Verified over 361 tasks with a 50-step budget, the best among 4B models in its table, against 68.7% for the 9B ScaleCUA-Qwen3.5 and 63.9% for Qwen3.6-27B.

## Why it matters locally

At 4B with weights under 10 GB and an Apache-2.0 license, this computer-use agent is a far smaller download than the 27B and 30B open models it scores alongside.

## For the market

Hosted frontier models still lead Microsoft's table, with GPT-5.5 at 78.7% and Claude Opus 4.7 at 78.0%, but the gap between a downloadable 4B model and GPT-5.4-mini at 69.6% is small. If these self-reported results hold up, routine desktop automation could become one of the first agent jobs to move on-premises.

## In plain words

This is a small AI that can use a computer like an assistant at your desk: it looks at the screen and clicks when it has to, and types commands when that is quicker. Because it is small and free to download, a company could run it on its own machine.

## For practitioners

Serve it with SGLang using Microsoft's Docker image or a transformers setup with trust_remote_code=True, since the llava_onevision2 architecture is custom; the sampling settings behind the 66.0% result are temperature 0.7, top_p 0.8 and 4,096 max new tokens. Microsoft publishes no minimum hardware, so measure memory with full 1920×1080 screenshots. The model can run arbitrary shell commands, so Microsoft's card says to always use an isolated container or VM; test on your own tasks too, since the benchmark used Linux desktops.

## Source

[microsoft/Mage-VL-Cowork model card (Hugging Face)](https://huggingface.co/microsoft/Mage-VL-Cowork)

## On this site

- Models: [Qwen3.5-9B](https://onpremworld.com/en/models/qwen3.5-9b.md)
- Models: [Qwen3.6-27B](https://onpremworld.com/en/models/qwen3.6-27b.md)

[All news](https://onpremworld.com/en/news.md)

Summaries are written by OnPrem World from the official source. We link to the original and never republish it.