Microsoft Open-Sources Mage-VL-Cowork, a 4B Computer-Use Agent

Microsoft's Mage team open-sourced Mage-VL-Cowork on Hugging Face under the Apache-2.0 license on October 30 — a 4B-parameter multimodal computer-use policy model. A single compact policy retains two action channels: mouse and keyboard operations on screen, plus a persistent shell (for writing code / running bash). At each step the model decides on its own which mode to use, with no extra router, addressing the inefficiency of traditional computer-use agents that only "watch screenshots and click the mouse".

Confirmed

• Released by Microsoft's Mage team, hosted on Hugging Face under Apache-2.0, with the official model card public
• Scores 66.0% on the OSWorld-Verified benchmark, reportedly surpassing 27B-parameter models per multiple posts — the best result at the 4B scale
• A hybrid computer-use agent combining computeruse and persistent shell action modes, chosen per step by the model itself, with no separate routing module

Why it matters

• A small 4B-scale vision-language model achieves competitive computer-use performance with much larger models for the first time, significantly lowering deployment barriers and costs
• The hybrid "clicking + coding" action design offers a new paradigm for GUI agent efficiency and may shape future computer-use agent architectures
• Apache-2.0 open sourcing means the community can directly reuse and build on it

2026-10-31 ~ 2026-10-31 · 9 related posts

• Microsoft releases 4B hybrid computer-use agent scoring 66% on OSWorld — _akhaliq · 2026-10-31
• [source] Microsoft releases Mage-VL-Cowork: a 4B desktop agent scoring 66.0% on OSWorld-Verified — pmttyji · 2026-10-31
• [source] Microsoft's Mage-VL-Cowork: a 4B agent that picks GUI clicks or bash each turn, 66% on OSWorld — AIFlow_ML · 2026-10-31
• Microsoft open-sources 4B agent that picks GUI clicks or bash each turn, scoring 66% on OSWorld — AIFlow_ML · 2026-10-31
• Microsoft open-sources Mage-VL-Cowork, a 4B agent scoring 66% on OSWorld-Verified — AIFlow_ML · 2026-10-31
• Microsoft open-sources 4B agent hitting 66% on OSWorld: cutting tool overlap before distilling adds 12.5 points — AIFlow_ML · 2026-10-31
• Microsoft open-sources a 4B GUI agent scoring 66% on OSWorld-Verified — AIFlow_ML · 2026-10-31
• Microsoft's open-source 4B agent picks clicks or bash per step, scoring 66% on OSWorld-Verified — AIFlow_ML · 2026-10-31
• Microsoft Open-Sources 4B Agent Choosing Clicks or Bash Each Turn, Scores 66% on OSWorld — AIFlow_ML · 2026-10-31