Mage-VL-Cowork
Microsoft published Mage-VL-Cowork, a 4B vision-language computer-use agent under Apache-2.0, claiming 66.0% on OSWorld-Verified with a 50-step budget.
One 4B model handles both GUI control (computer_use) and terminal execution (bash). It was trained in five stages: GUI midtraining, GUI agent SFT, coding and terminal SFT, hybrid computer-use SFT, and online reinforcement learning with GRPO. Microsoft says online RL cut output tokens by 25.4% and model calls by 13% compared with SFT alone.
- Date
- Friday 30 October 2026
- Lab
- Microsoft
- Kind
- open-weights
- Access
- open weights
Figures
| Measure | Value | Measured by |
|---|---|---|
| OSWorld-Verified | 66.0% 50-step budget | company |
| Operating System tasks | 87.5% | company |
| Multi-Application Workflows | 62.7% | company |
| ScreenSpot-v2 | 87.2% | company |
All benchmark scores are reported by Microsoft and not independently verified.
Sources
- huggingface.co/microsoft/Mage-VL-Cowork
- pasqualepillitteri.it/en/news/28914/mage-vl-cowork-microsoft-4b-computer-use-agent-osworld
This record was checked and corrected against its sources on 1 November 2026. How we check