Microsoft open-sources 4B agent hitting 66% on OSWorld: cutting tool overlap before distilling adds 12.5 points
AIFlow_ML · x · 2026-10-31
Microsoft open-sourced a 4B agent model that picks between mouse clicks and bash each turn without a router, scoring 66% on OSWorld-Verified. Key lesson from the model card: pruning overlapping tools before distillation gave the 4B student +12.5 points — more tools can hurt small models. Deployers should sandbox bash execution first.
Related event: Microsoft Open-Sources Mage-VL-Cowork, a 4B Computer-Use Agent(9 posts)→
More from coding & agent
- Dev shares safety architecture for self-improving agent: out-of-band release controller with passkey approvals — viksit · 2026-11-01
- Dev shares Grok Bot team-building guide: few focused bots that write their own skills — CurieuxExplorer · 2026-11-01
- Vercel Labs open-sources agent-desktop: computer use for any coding agent on macOS, Windows, Linux — evilrabbit_ · 2026-11-01
- 747 AI-agent ideas reveal a scaling law for idea descent: +17% per doubling — DimitrisPapail · 2026-11-01
- Magnific Move + Claude Opus: Halloween game animated in one chat — petewoodbridge · 2026-11-01
- AI-Driven Frame Rate Optimization Cuts Render Time by 9%-15% Across Six Scenes — AIandDesign · 2026-11-01