Mage-VL-Cowork

Mage-VL-Cowork is a 4-billion-parameter computer-use agent (CUA) released by Microsoft's Mage team on October 30, 2026, with open weights on Hugging Face under Apache-2.0. Its tagline is "One 4B policy that knows when to click — and when to code."

Unlike agent systems that route between a separate GUI operator and a coding operator, Mage-VL-Cowork is a single vision-language policy that sees full-desktop screenshots and, turn by turn, decides whether to emit mouse/keyboard actions or run commands in a persistent shell/Python environment. Microsoft reports 66.0% on OSWorld-Verified under a 50-step budget — the best result it knows of at the 4B scale, and above larger open agents such as Muse Glimmer-30B (65.9%) and Qwen3.6-27B (63.9%).

Key points

1. Hybrid action space: two channels — computer_use (11 GUI primitives on normalized coordinates) and bash (persistent shell + Python). The policy picks the channel itself; no external router or multi-agent orchestrator.
2. Training recipe: five-stage progressive curriculum (midtraining → GUI SFT → hybrid SFT → online RL) starting from an intermediate Mage-VL checkpoint. Scores climb 45.56% → 55.84% → 66.0% across stages; RL adds +10.16 pp while cutting model calls 13% and output tokens 25%.
3. Mage-CUA-Forge: a workflow-first task synthesizer that generated 293,806 desktop/web tasks across 6,915 skills, reducing per-task cost by 82.6%.
4. Results: 87.5% on OSWorld OS tasks, 62.7% on multi-app workflows; big gains on ScreenSpot-v2 and OSWorld-G grounding.
5. Caveats: numbers are self-reported with Microsoft's own MAGEHybridAgent harness; independent reproductions are pending.

Why it matters

A 4B model is small enough to run on a single consumer GPU or high-end laptop, so Mage-VL-Cowork pushes capable desktop automation toward on-device use — while also raising the usual safety concerns of letting a model run shell commands on a real machine.

How to try it

1. Read the project page and the technical report (PDF linked there).
2. Download weights from Hugging Face (microsoft/Mage-VL-Cowork) and use the provided MAGEHybridAgent harness to reproduce the OSWorld setup (1920×1080, three screenshots of history).
3. Run it inside a VM or sandbox first — it can execute bash on the machine it controls, so add permission and policy layers before touching real data.
4. Good fits: office-app automation (spreadsheets, mail, file management), QA of desktop software, and research on small on-device agents.