Microsoft Mage-VL-Cowork Is a 4B Computer-Use Model That Knows When to Click and When to Type

Microsoft has posted Mage-VL-Cowork, a 4 billion parameter model for using a computer, and the idea behind it is easy to like. Most computer-use agents work only from screenshots. They look at the screen, move the mouse, click, and look again. Mage-VL-Cowork can do that, but it also has a terminal, and it decides for itself which one to use at each step.

Click or type

According to the model card, the model uses the mouse and screenshots for visual jobs like layouts, canvas editing, or an app's live state. It switches to bash for structured edits to spreadsheets, documents, and slides, and for checking its own work. There is no separate router or second agent making that call. It is one model choosing at each turn.

That is roughly how a confident person uses a PC. You click around a photo editor, but you don't rename two hundred files by hand.

The numbers, with a caveat

Microsoft reports 66.0 percent on OSWorld-Verified with a 50-step limit. That is the average of two full runs over 361 tasks, and the card calls it the best score at the 4B size. It lists 87.5 percent on operating system tasks and 62.7 percent on workflows that span several apps. Large proprietary models still score higher in the same table, and all of these results are Microsoft's own.

The training ran in five stages, ending with online reinforcement learning. Microsoft says that last stage added about ten points while cutting output tokens per task by roughly a quarter. To save memory, the model keeps only the three latest screenshots and folds older steps into one-line summaries.

Can you run it

Yes. The weights are on Hugging Face under the Apache 2.0 license, at about 9.5 GB in bfloat16. Community quantized versions showed up within a day. A small agent like this is what could make on-device helpers practical, and I'm curious how it handles messy everyday apps.

Details are on the official model card.