Microsoft's 4B Mage-VL-Cowork hits 66% on OSWorld-Verified SOTA; Niels Rogge builds runnable demo on HF Sandboxes
NielsRogge · x · 2026-11-04
Niels Rogge demoed Microsoft's new 4B model Mage-VL-Cowork, which sets a new SOTA of 66% on OSWorld-Verified (excluding Google Drive tasks) for on-device computer use, beating much larger models.
Key details:
- Hybrid agent: at each step it decides whether to use the GUI (clicks, keys) or a Bash shell; for structured edits it often just writes Python instead of clicking through menus
- Demo architecture: model runs in an HF Space, desktop in an HF Sandbox (isolated VM booting in seconds, Xvfb + XFCE + Firefox + xdotool image, ready in 18s)
- Loop: screenshot via sandbox file API → model returns think + action + tool calls → xdotool executes clicks/typing, bash runs in the same box; Gradio UI streams the screen, reasoning, and a red dot where it clicked
A compelling lightweight-model + hybrid-tooling path for on-device agents, with a fully reproducible HF Sandbox setup.
Related event: Microsoft's 4B Mage-VL-Cowork Sets On-Device Computer Use SOTA on OSWorld(3 posts)→
More from coding & agent
- Phone Harness (YC F26) open-sources a harness that lets AI agents control real phones — ycombinator · 2026-11-05
- Inside a16z's agent podcast: trust is the top fear among 147 tracked AI assistants — a16z · 2026-11-05
- The vibe coding blind spot: people can't articulate their AI-built products' benefits — rms80 · 2026-11-05
- Non-dev built an agent-only eSIM store; a Claude agent bought one with real USDC — Apprehensive_Big5042 · 2026-11-05
- Google Cloud's five-rule playbook for evaluating whether agent skills are worth their tokens — rseroter · 2026-11-05
- Google DeepMind 工程师将登 AIE CODE 峰会,讲述删掉 90% 评估框架的故事 — _philschmid · 2026-11-05