- Microsoft introduces Mu, a model of IA designed to run locally on PCs with Windows 11 and Copilot+.
- Mu is lightweight (330 million parameters), fast (over 100 tokens/second), and optimized for NPUs.
- Allows you to perform actions in Settings using natural language and without a cloud connection.
- Trained with big data and distillation technologies for high performance and efficiency.

In the last days Microsoft has caused a stir in the world of Artificial Intelligence when presenting Mu, its new, compact language model. This advance It is designed to run directly on computers with Windows 11, specifically those that have Copilot+ technology. The objective of this development is clear: to offer a Fast and efficient performance from the device itself, without having to rely on a cloud connection.
With the introduction of Mu, Microsoft bets on local and lightweight AI models, capable of executing tasks directly on the Neural Processing Unit (NPU) of the most advanced PCs. This initiative responds to a growing trend in the sector, where other companies are also seeking to reduce their dependence on cloud services and provide their devices with greater autonomy and immediate response.
Key Features of Mu: Agile and Efficient AI for Windows 11

The great attraction of Mu lies in its small size and its optimization to operate locally.. It has 330 billion parameters, allowing it to install itself completely on the computer's NPU and respond with outstanding speed (more than one hundred tokens per second)This figure surpasses the company's previous models, while maintaining a fluid, wait-free user experience.
The uses an encoder-decoder architecture, which means that it transforms the natural language entered by the user into an internal representation that is then converted into the desired response. In this way, reduces memory and processing load and you get higher performance in hardware and, very useful for everyday tasks with limited resources.
Mu has been trained with a huge educational dataset and distillation techniques derived from other models such as Phi. This allows it to understand complex queries, learn grammar and semantics, and provide precise instructions when modifying operating system settings.
Among the technological innovations it incorporates are:
- Grouped-Query Attention (GQA), which reduces memory footprint without sacrificing precision;
- Rotary Positional Embeddings (RoPE), for better interpretation of long sequences;
- Dual LayerNorm, stabilizing training and improving its adaptability.
Mu as the Configuration Agent Engine on Windows
El The first major practical use of Mu is in the Windows 11 Settings agent.. On Copilot+ PCs, this model allows users to interact in natural language with the operating system to adjust options without having to manually navigate through menus. This way, you can simply type or say actions like “put on dark mode” o “Show me the available Wi-Fi networks” for the AI to make those changes immediately.
Thanks to its compact size and its complete download process at the NPU, the model can respond to requests in less than half a second. Microsoft claims that Mu's speed and efficiency are comparable to models ten times larger, such as Phi-3.5, which represents a significant leap in the integration of local AI for consumers.
It's worth noting that engineers trained Mu with millions of examples of real-world instructions and adapted the model to solve both general and specific queries. In the case of ambiguous requests, the system can resort to traditional methods such as keyword search.
Mu Training, Optimization, and Availability
For the development and training of Mu, Microsoft has deployed powerful GPUs in the Azure cloud, applying advanced distillation techniques and low-rank adaptation to fine-tune the model to the specific needs of the Configuration Agent. Special emphasis has been placed on cover hundreds of possible actions with synthetic and natural data, improving the model's ability to understand instructions as any user would give them.
By distributing the load between the capabilities of the NPU and the algorithmic efficiency of Mu, Energy and memory consumption is minimal, promoting better autonomy and performance in computers portable, something key for the Copilot+ PC users.
At present, Mu is available to users who are part of the Windows Insider Program, although Microsoft is expected to expand access to other audiences as it continues to refine the model and receives feedback on actual usage.
The launch of Mu represents a significant step forward in Microsoft's strategy to equip its systems with local AI, optimized for specific hardware and designed for everyday tasks. Integration with the Configuration agent may be just the beginning of a new generation of assistants capable of understanding and executing commands directly from the computer itself, without relying on remote servers. This places Microsoft in a prominent position in the race for the next evolution of artificial intelligence applied to personal computing.
Passionate writer about the world of bytes and technology in general. I love sharing my knowledge through writing, and that's what I'll do on this blog, show you all the most interesting things about gadgets, software, hardware, tech trends, and more. My goal is to help you navigate the digital world in a simple and entertaining way.