Microsoft Releases Mage-VL-Cowork: An Efficient 4B Desktop Agent Model
Microsoft has introduced Mage-VL-Cowork, an efficient 4B small vision-language model that achieves state-of-the-art performance on the OSWorld-Verified benchmark for autonomous desktop agents. The model combines upstream teacher-harness evolution, workflow-first environment synthesis, and a specialized curriculum to master hybrid computer tasks. This release pushes the frontier of on-device intelligence by proving that compact 4B models can achieve top-tier desktop automation performance while significantly reducing computational overhead. It enables more responsive, cost-effective local AI agents that can interact with desktop interfaces. Mage-VL-Cowork achieved 66.0% accuracy on OSWorld-Verified while using 13.0% fewer model calls and 25.4% fewer output tokens after online reinforcement learning. It handles hybrid computer use by integrating mouse and keyboard inputs for visual layouts with bash commands for structured documents.
## BACKGROUND
OSWorld-Verified is a rigorous benchmarking environment used to evaluate multimodal AI agents on open-ended, real-world computer tasks. Small vision-language models (VLMs) like this 4B parameter model are designed to process both visual screen information and textual instructions directly on consumer hardware.