Mage-VL-Cowork has two tools in one policy: computer_use for mouse and keyboard on screenshots, and bash for shell and Python. It picks per turn, so a spreadsheet edit can go through openpyxl in the terminal while an image edit happens in the GIMP window. It's built on Qwen3-4B with Microsoft's Mage-ViT vision encoder and runs a 32,768-token context.
Microsoft's figures: 66.0% on OSWorld-Verified (the average of two full runs, 361 tasks excluding Google Drive, 50 steps), against 68.7% for a 9B ScaleCUA model and 69.6% for GPT-5.4-mini in its table. In one demo it finished a LibreOffice Calc task in 39 seconds and 6 steps, against 9 minutes 56 seconds for Qwen3.6-27B clicking through the GUI. These are Microsoft's own numbers.
It needs trust_remote_code because the architecture is custom, and Microsoft ships a pinned SGLang Docker image. The model card says to run it only in a disposable container or VM with no host files, keys or cookies mounted.

