Microsoft's Mage team published Mage-VL-Cowork on Hugging Face as an open Apache-2.0 vision-language agent for computer use. At about four billion parameters, a single policy can click and type from screenshots through a computer_use channel, or run shell and Python commands through bash, without an external router choosing the channel.
On OSWorld-Verified with a fifty-step budget, the model card reports 66.0 percent overall, among the strongest published results at the 4B scale. Category scores include about 87.5 percent on operating-system tasks and 62.7 percent on multi-application workflows. The checkpoint ships as two safetensors shards totaling roughly 9.48 GB in bfloat16.
Training follows a five-stage specialize-then-unify curriculum: GUI grounding and agent fine-tuning, coding and terminal fine-tuning, hybrid supervised fine-tuning, then online reinforcement learning. Upstream, Mage-CUA-Forge synthesizes about 293,806 desktop and web tasks across 6,915 atomic skills with reusable workflows, cutting recurring per-task synthesis cost by a large margin according to the authors.
For people who already run local coding agents, Mage-VL-Cowork shows that a compact open model can both drive the desktop UI and edit files programmatically. Official examples include merging spreadsheet cells with Python and verifying image edits from the terminal after GUI steps. It is a useful comparison point if you ship websites or automations with Claude Code-style workflows on your own machine.