Microsoft’s Mage team released Mage-VL-Cowork, a 4B “hybrid” computer-use agent. One vision-language policy can either click and type in a GUI from screenshots or drop to code and the terminal, choosing the channel each turn without an outside orchestrator. The brief lists it as Apache-2.0.
Numbers
- 66.0% on OSWorld-Verified with a 50-step budget
- Ahead of bigger open agents like Qwen3.6-27B (63.9%) and OpenCUA-72B (44.9%); ScaleCUA-Qwen3.5-9B is higher overall at 68.7%, but Mage leads it on OS tasks (87.5% vs 79.2%) and multi-app workflows (62.7% vs 48.1%)
- Frontier models like GPT-5.5 (78.7%) are still ahead
The team also built Mage-CUA-Forge, which generated 293,806 desktop and web tasks across 6,915 skills to train it.
My take
A 4B model that can drive a desktop is small enough to run locally, which is exciting and a little scary. An agent that clicks and runs commands can wreck things fast.
Review before you install: read the release notes yourself, back up your config, and try new versions somewhere you can afford to break first.
Isolate it: if you try this, run it inside a disposable VM with no access to your real accounts, keys or files, and snapshot before you start.
God Bless, and Happy hacking!
