Microsoft open-sources a 4B GUI agent scoring 66% on OSWorld-Verified
AIFlow_ML · x · 2026-10-31
Microsoft open-sourced a 4B agent that decides each turn between mouse clicks and bash commands, with no separate router. Its card reports 66.0% on OSWorld-Verified, edging out Qwen3.6-27B's 63.9% — company numbers, so verify independently. Note: bash support means you'll want a sandbox for deployment.
Related event: Microsoft Open-Sources Mage-VL-Cowork, a 4B Computer-Use Agent(9 posts)→
More from coding & agent
- The hard part of coding agents isn't the code — it's the 4 jobs around every PR — Arindam_1729 · 2026-10-31
- Your LLM never sees your prompt — it sees tokens: a tokenization primer — mdancho84 · 2026-10-31
- Track Codex/Cursor agent tasks live via macOS Live Activities — Dimillian · 2026-10-31
- LangChain's Harrison Chase: context engineering is far from solved — here's what's still wide open — hwchase17 · 2026-10-31
- Weekend watch: 2 hours on post-training agents with open models — SergioPaniego · 2026-10-31
- TUI tools embrace floating panes: tuios shows off multi-agent terminal OS — sull · 2026-10-31