Microsoft Mu: A New Era of On-Device AI for Windows 11

Microsoft is ushering in a new phase of user interaction with the introduction of Mu, a small language model (SLM) designed to run directly on Windows 11 devices. This on-device AI aims to revolutionize how users manage system settings, offering a more intuitive and private experience.

First unveiled in June 2025, Microsoft Mu is a compact and efficient AI model that processes natural language queries to control various Windows 11 settings. Unlike cloud-based AI, Mu operates locally on the device, ensuring that user commands are executed quickly and privately. This development marks a significant step towards more personalized and secure computing, leveraging the power of dedicated Neural Processing Units (NPUs) in modern PCs.

Streamlining User Experience with Natural Language

At its core, Mu is designed to simplify the often-complex process of navigating Windows settings. Instead of clicking through multiple menus, users can employ simple, conversational commands to adjust their system's configuration. For example, phrases like "lower the screen brightness" or "how do I turn on night mode?" can be directly understood and acted upon by Mu.

This functionality is integrated into the search box within the Windows 11 Settings app, providing a seamless experience for the user. With the ability to interpret and manipulate hundreds of system settings, Mu can even take action on behalf of the user with their permission.

The Power of On-Device Processing and NPUs

A key characteristic of Microsoft Mu is its on-device processing capabilities. This means that all AI computations happen locally, eliminating the need to send data to the cloud. This approach offers several advantages:

  • Enhanced Privacy: User commands and data remain on the device, significantly improving privacy and security.
  • Increased Speed: Local processing allows for near-instantaneous responses, with Microsoft claiming response times of under 500 milliseconds.
  • Offline Functionality: As the model runs locally, it can function without a constant internet connection.

To achieve this, Mu is optimized to run on the Neural Processing Units (NPUs) found in the latest Copilot+ PCs. This specialized hardware is designed to handle AI workloads efficiently, enabling Mu to operate at speeds of over 100 tokens per second. While initially exclusive to devices with Qualcomm's Snapdragon X chips, Microsoft has plans to expand support to AMD and Intel-based PCs in the future.

Technical Advancements and a Sibling to Phi Silica

Microsoft describes Mu as a highly efficient model with 330 million parameters, optimized for on-device deployment. It is built on a transformer encoder-decoder architecture, which contributes to its efficiency by separating input and output tokens, thereby reducing computational and memory overhead.

Mu's development builds on Microsoft's experience with its Phi family of small language models. It is considered a highly optimized "sibling" to the Phi Silica model, which is also used in Windows 11 for Copilot+ PC features. Through various optimization techniques, Mu is able to deliver performance comparable to the much larger Phi-3.5-mini model while occupying ten times less space. The initial training of Mu was conducted on NVIDIA A100 GPUs on the Azure Machine Learning platform.

Availability

Microsoft Mu is currently available in preview for Windows Insiders in the Dev Channel who are using Snapdragon-powered Copilot+ PCs. A wider release for other compatible hardware is expected at a later, unspecified date. This phased rollout allows for further refinement and optimization of the technology before it becomes a standard feature in Windows 11.