AI Briefing
KOSign in

Microsoft Releases Mage-VL-Cowork: 4B VLM Achieves 66.0% on OSWorld-Verified

·2026.10.31 19:02

Key point

The 4B model uses 13.0% fewer model calls and 25.4% fewer output tokens than comparable systems after online RL.

Details

  • Microsoft introduces Mage-VL-Cowork, a 4B vision-language model designed for hybrid computer use and autonomous desktop agents.
  • The model achieves 66.0% on OSWorld-Verified, the best performance at the 4B scale.
  • Efficiency gains include 13.0% fewer model calls and 25.4% fewer output tokens following online RL.
  • Performance relies on upstream teacher-harness evolution, workflow-first environment synthesis via Mage-CUA-Forge, and a five-stage specialize-then-unify curriculum.
microsoftvision-language-modelcomputer-useefficiencymage-vl-coworkosworld-verified
Read originalShare on Threads

This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.

Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.

Related

Hacker NewsOpenAI releases GPT-6.1 Terra with 50% lower API costs and improved computer useRedditMicrosoft Q1 revenue hits $92.3 billion as Azure cloud growth accelerates to 48%RedditLIMe Reduces LLM Training Compute by 15% Without Quality LossTLDR AIVisual Studio October Update Adds Commit-to-Working-Tree Comparison and Bulk Conflict ResolutionPyTorchKRMicrosoft and ThinkingBox-Training Released: Qwen3.8-27B Repeat Success Rate Improved 2.8xVercel BlogOpenAI GPT-6.1 Terra Available on AI Gateway with 1M+ Token Context