Microsoft has unveiled a new compact language model that runs entirely on the NPU of Copilot+ PCs.
The new language model called Mu forms the brain of the AI agents now being tested in the Windows Settings app. Microsoft emphasizes that it runs entirely on the NPU “s of Copilot+ PC” s.
What Can MU Do?
Mu was specifically designed to fit within the memory and energy constraints of NPU “s”. Thanks to this optimization, the agent responds lightning-fast. Windows Central states that if you type, for example, “my mouse cursor is too small”, a suggestion to enlarge the pointer appears immediately. The user no longer needs to search through sub-menus; the agent executes the change as soon as permission is granted.
Microsoft engineers reported that existing models were too slow for the desired latency. Therefore, they refined Mu with hundreds of practical scenarios. The result is a latency of less than half a second for hundreds of settings, making the interaction feel natural.
Within the Boundaries of the NPU
Vivek Pradeep, Corporate Vice President for Windows Applied Sciences, emphasizes that Mu uses NPU-optimized operations. This allows the model to avoid unnecessary processor switching and keeps system load low, even with repeated use.
Microsoft is now rolling out the feature in the Dev channel of the Insider program. If the feedback is positive, Mu will also appear outside the test program, allowing all Copilot+ users to adjust settings using natural language.