Microsoft's Mage-VL-Cowork: a 4B agent that picks GUI clicks or bash each turn, 66% on OSWorld
AIFlow_ML · x · 2026-10-31
Microsoft's Mage team published Mage-VL-Cowork on Hugging Face (Oct 30, Apache-2.0), a 4B multimodal policy fixing the inefficiency of computer-use agents that only click: tasks like merging spreadsheet cells burn dozens of clicks when one Python call would do.
Design
- Two action channels in one model: computeruse (mouse/keyboard) and bash (persistent shell)
- No external router — the model picks the channel each turn
- 4B params, Apache-2.0
Scores (OSWorld-Verified, 50-step, vendor-reported)
- Human ref: 72.4%; GPT-5.4-mini: 69.6%
- Mage-VL-Cowork 4B: 66.0%, above Muse Glimmer 30B (65.9%) and Qwen3.6-27B (63.9%); Microsoft's own Fara1.5-4B only 13.9%
The author argues the bash channel is the real leap — and the reason you must sandbox it before running.
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
- 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
- Parcha founder: You may not need an agent after all — decision models vs workflows — blaizedsouza · 2026-11-01