Skip to content
AI Model Radar

Mage VL Cowork

microsoft/Mage-VL-Cowork
Open Source · chat · open-weights
Open Alert me on changes
Context
32.8K
Max output
—
Weights
Open
API $/1M
$0.04 / Free
Modalities
text · image
Released
30 Oct 2026
License: apache-2.0 · microsoft/Mage-VL-Cowork
Download image Share on X Share on LinkedIn
AI summary
● machine-written

Microsoft releases Mage-VL-Cowork, 4B parameter vision-language desktop agent

Mage-VL-Cowork is a 4-billion parameter open-weights vision-language model designed for computer use tasks on desktop environments. The model achieves 66.0% performance on OSWorld-Verified benchmarks under a 50-step budget, positioning it as the top performer among 4B-class models for desktop automation and on-device intelligence.

What's new
  • 4B parameters optimized for desktop agent tasks with 32,768-token active context window
  • 66.0% OSWorld-Verified performance, best among 4B models
  • Online reinforcement learning training adds +10.16 percentage points over Hybrid SFT while reducing output tokens by 25.4%
  • Quantized Q4 version fits in 2.7GB memory on 16GB systems
Best for
On-device desktop automation and computer use tasksResource-constrained environments requiring efficient VLMsUI navigation and desktop agent applications
Sources

Source: https://huggingface.co/microsoft/Mage-VL-Cowork