Microsoft Open-Sources Mage-VL-Cowork, a 4B Hybrid Computer-Use Agent
Microsoft Mage team open-sources Mage-VL-Cowork, a ~4B hybrid agent switching GUI clicks and terminal code in one policy; ~66% OSWorld-Verified, Apache-2.0.
Computer-use agents have long faced an awkward fork: stare at pixels and click, or split GUI work and coding across systems glued together by an orchestrator. Microsoft's Mage team chose a cleaner answer—one roughly 4B-parameter policy that decides, turn by turn, whether to move the mouse or open a terminal.
The model is Mage-VL-Cowork. Weights are on Hugging Face, with a project page and technical report, under Apache-2.0. Training starts from an intermediate Mage-VL checkpoint and follows a five-stage specialize-then-unify curriculum: GUI-oriented midtraining, GUI agent SFT, coding and terminal SFT, hybrid CUA SFT, and online hybrid reinforcement learning. The bet is not more parameters, but teaching the model when to click and when to code inside a single policy.
On OSWorld-Verified with a 50-step budget, Mage-VL-Cowork reaches about 66.0%—a two-run average on 361 held-out tasks excluding Google Drive—strong at the 4B scale and ahead of several larger open computer-use agents. Category scores include roughly 87.5% on OS tasks and 62.7% on multi-application workflows. Online RL adds about 10 points over hybrid SFT while cutting mean model calls and output tokens. A small model that is both more accurate and more compact is the real headline.
Two upstream pieces matter as much as the checkpoint. Teacher harnesses are evolved before trajectory collection, so redundant tool habits are not distilled into the student. Mage-CUA-Forge synthesizes about 293,000 desktop and web tasks with skill-conditioned workflows, cutting recurring per-task generation cost by more than 80%. In demos, spreadsheet edits go through Python in the terminal rather than dozens of GUI drags—the hybrid move real desktop automation actually needs.
A 4B open agent that can touch a full desktop is already an argument for on-device agents. OSWorld is still a sandbox score, and production pain—app version drift, permission dialogs, long-horizon recovery—will bite later. But the direction is clear: computer use no longer has to choose between pure GUI and pure code. Microsoft put the switch inside one small policy.