REDMOND, Wash. – Microsoft has released an open-weight artificial intelligence model designed to work across desktop applications, switching between clicking on the screen and using code to complete tasks.
The company’s Mage team published Mage-VL-Cowork on Friday, October 30, with model weights, documentation and reference software for evaluating its performance. The release is aimed at developers and researchers, rather than being a new feature inside a consumer version of Windows.
Microsoft describes the model as a compact, 4-billion-parameter computer-use system. It can interpret screenshots and choose mouse or keyboard actions, while also using a terminal to make structured file changes. That allows it to use a visual interface for a task such as editing an image, or code for changes to a spreadsheet or document.
The team reports a 66.0% score on OSWorld-Verified, a benchmark of desktop tasks. That is Microsoft’s own reported result, not an independent test by JBizNews.
Its evaluation covered 361 tasks, excluded Google Drive and allowed up to 50 steps per task. The headline result is an average of two full runs. Those conditions matter when comparing it with scores reported for other systems.
The release makes the model’s weights available under the Apache 2.0 license. Microsoft’s documentation also describes a five-stage training process that first develops visual-interface and coding skills separately, then trains the model to choose between them.
The ability to act on a computer also creates risks. The model card recommends isolated test environments and human confirmation before irreversible actions such as financial transactions or changes to account privileges. It warns against giving the system access to personal credentials or files without a specific, limited need.
The release materials do not establish that the model can safely handle every real-world desktop task without supervision. They provide developers with a downloadable model and a reported research benchmark, rather than a guarantee of reliable autonomous office work.
The model and documentation are available on Microsoft’s Hugging Face page and its official GitHub repository.
JBizNews Tech Desk | Redmond, Wash.
© JBizNews.com. All rights reserved. Unauthorized reproduction or redistribution is strictly prohibited.
