Microsoft releases Mage-VL-Cowork, a 4B computer-use agent at 66% on OSWorld

Microsoft has released Mage-VL-Cowork, a computer-use agent with just 4 billion parameters that scores 66% on OSWorld-Verified, according to coverage from Pasquale Pillitteri and XenoSpectrum on October 31.

The two numbers in the headline are the story. Four billion parameters is a small model, and 66% is a strong result on a benchmark built to test desktop agents end to end. The coverage describes it as the best score at the 4B scale.

A computer-use agent drives software the way a person does: it reads the screen, moves the cursor, and types, rather than calling a tidy API. The usual pattern is to put a frontier model behind the wheel, which is expensive, slow, and hard to run on a local device.

OSWorld is the benchmark aimed at exactly that skill. It asks an agent to complete multi-step tasks inside real desktop environments, covering system settings, office applications, and workflows that cross between programs. A 66% score means the model finishes a substantial share of those task chains on its own.

Microsoft also published the weights, the code, and a technical report. The Hugging Face model card lists an Apache 2.0 license, so teams can download it, deploy it, and run their own evaluations instead of judging from a demo video.

That combination shifts the competitive story. Progress over the past year has been read in parameter counts, but a 4B model that is roughly usable on desktop work suggests the next contest is about how many tasks each dollar buys and whether the thing runs on ordinary hardware.

Three things to watch: whether outside teams reproduce the 66% figure, how stable the agent is in messy real-world enterprise desktops, and whether 4B holds up when a task runs for dozens of steps rather than a handful.