Microsoft is moving natural language control deeper into Windows with a new AI assistant for the Settings menu. The feature lets users ask for changes or describe problems in plain English, instead of manually finding the right page, toggle, or slider.
The system is powered by a new language model called Mu. It runs offline on specialized AI chips and is currently limited to a small group of users: Windows Insiders using Copilot+ PCs in the Dev Channel.
What the Windows Settings AI agent does
The assistant is built directly into the existing Settings search box. That placement matters because it keeps the feature close to the workflow people already use when they do not know exactly where a Windows option lives.
Instead of navigating through menus, a user can type a request such as "Increase the screen brightness." The system can also respond to problem descriptions, including "My mouse pointer is too small." In both cases, the goal is to translate ordinary language into the relevant Windows settings action.
According to Microsoft, the assistant responds in less than half a second. That response target is important for a settings tool because it has to feel closer to search or direct control than to a long chatbot exchange.
The feature is not meant to replace every settings interaction at once. Microsoft is currently focused on the most commonly used settings, while support for more complex tasks is planned for the future.
Why Mu is built differently
Mu is a purpose-built language model rather than a general assistant placed inside Windows Settings. Microsoft developed it specifically for the task of understanding settings requests and producing the appropriate Windows commands.
The model has 330 million parameters. It uses an encoder-decoder architecture, which Microsoft says is much more efficient than the decoder-only models commonly used today.
In this design, the encoder reads the user request once and converts it into an internal representation. The decoder then uses that representation to generate the Windows command needed for the requested action.
Microsoft says this split reduces response times by 47 percent and increases processing speed by a factor of 4.7 compared to similar models. The result is a smaller, more targeted system that is optimized for one narrow job rather than open-ended conversation.
Local AI depends on Copilot+ PC hardware
Mu runs entirely on the Neural Processing Units, or NPUs, in Copilot+ PCs. These dedicated AI chips are designed for machine learning workloads, allowing the settings assistant to work locally instead of relying on cloud processing.
Microsoft says Mu can process over 100 tokens per second on NPU hardware. On a Surface Laptop 7, the system can reach more than 200 tokens per second.
To make the model work efficiently on this hardware, Microsoft used post-training quantization. That process converts model weights from floating-point values to 8-bit and 16-bit integers.
The practical effect is lower memory use and higher processing speed. For a feature inside the Settings menu, those gains support the broader design goal: fast local responses for focused system-control tasks.
How Microsoft trained the settings agent
Microsoft began by training Mu with hundreds of billions of educational tokens. It then distilled knowledge from the company's larger but still small Phi models.
For the Windows Settings agent itself, Microsoft used 3.6 million training examples, including some synthetic examples. That expanded coverage from about 50 Windows settings to hundreds of Windows settings.
The assistant works best when the user gives a longer and clearer request. Short or unclear commands may still lead the search box to show ordinary search results instead of taking action.
One example is brightness control. Microsoft notes that "Increase brightness" can create issues when several monitors are connected. A clearer request gives the system more context and reduces ambiguity.
What limited availability means
The new Windows Settings AI agent is not broadly available yet. Microsoft is collecting feedback from Windows Insiders, and there is no word yet on when the feature will roll out to all Windows users.
Hardware will also limit early adoption. Because Mu depends on NPU hardware found only in the latest Copilot+ PCs, many Windows users will not be able to try it immediately.
The feature points to a larger direction for personal computing: operating systems that can be controlled through natural language. Microsoft has previously demonstrated experimental agent systems that interact with GUIs through multimodal language models.
Mu takes a more specialized path. Training a model for a very specific task may make it less flexible, but Microsoft is showing that this narrower approach can be far more efficient for a practical system feature.