Microsoft releases Mage-VL-Cowork, a 4B computer-use agent it reports at 66.0% on OSWorld-Verified

Teaser chart comparing Mage-VL-Cowork's OSWorld-Verified score with larger computer-use agents — Microsoft

The project page says the open model combines GUI and terminal actions in one policy, scores 66.0% under a 50-step budget ahead of larger agents including Qwen3.6-27B, and that online RL added 10.16 points; the results are Microsoft's own.

Read at Microsoft Mage

+ Hugging Face model card

More from Snape Report

Meta built a humanoid robot prototype with China's Fourier and is working on a second version, The Information reports
Anthropic finds Claude can write training data that passes a production-style monitor yet makes a model misaligned
Northflank says customers' AI API keys were abused after a breach, one of nine AI platform incidents Wiz counts
Cisco Talos details VOLTA infostealer that asks Claude, Kimi or Gemini to rank what each victim is worth
DeepSeek publishes its Super-agent design and open-sources dsh-superagent, a proactive coding-agent prototype
OpenAI has about $11 billion committed for its round, less than half its first $30 billion target, WSJ reports
Nasdaq slides 2.9% as Intel has its worst day since the 1970s and chip stocks fall
Apple confirms Apple Intelligence launches in China in December with iOS 27.3, with Alibaba's Qwen

Front page →