Windows ist nun mit dem von Microsoft eigens entwickelten Modell Mu ausgestattet. Mit einem einzigen Satz können alle Systemeinstellungen vorgenommen werden.

Still struggling with complex Windows settings? Microsoft is redefining the interaction of the settings interface.

The newly released on-device small language model Mu enables the settings in Windows 11 to have its own AI Agent.

With it, finding and changing PC settings becomes much easier. Now, all you need is a simple question, for example:

My mouse pointer is too small.

Mu will immediately recommend solution steps. Click "Apply" with one click to automatically execute the task and get a Plus version pointer.

Or do you want to control your computer by voice?

You can also directly enter your needs in the search bar to complete the settings.

Now you don't have to ask Baidu or Xiaohongshu everywhere about how to set a certain function. Those who understand will be in tears TT.

Mu is targeted at Copilot+PC and can directly map natural language queries to settings function calls. It runs locally quite efficiently, providing over 100 tokens per second.

Its performance is comparable to Microsoft's proud Phi model, but it is only one-tenth the size of the Phi model, about 330M.

This function can be found in the query section of the "Settings" menu in the Windows 11 preview version. It requires a Copilot+ computer powered by Snapdragon. Later, it will be extended to PCs supported by AMD and Intel™.

Kind reminder: Currently, only English conversations are supported.

Enable the Agent in Settings

Mu is an efficient 330M encoder-decoder model specially optimized for small-scale deployment.

The model is built based on the Transformer architecture, which means the encoder will convert the input into a fixed-length latent representation, and then the decoder will generate output tokens.

By separating input tokens and output tokens, Mu's one-time encoding significantly reduces computing and memory requirements, with lower latency and higher throughput.

Like Phi-Silica, Mu is designed to run efficiently on NPU and completes NPU adaptation by fine-tuning the model architecture and parameter shapes:

Model architecture adjustment: Select layer dimensions (such as hidden layer size and feed-forward network width) that match the tensor size and vectorization unit preferred by the NPU to ensure that operations such as matrix multiplication can run at peak efficiency.

Parameter shape change: Allocate parameters between the encoder and decoder in a 2/3 - 1/3 ratio. For example, a configuration includes 32 encoder layers and 12 decoder layers to ensure maximum performance per unit parameter.

To reduce the total number of parameters, Mu adopts the weight sharing method in some components, saving memory space and improving the consistency of the encoding and decoding vocabularies.

In addition, Mu limits operations to only run supported NPU optimization operators, avoiding invalid operations and fully utilizing the acceleration capabilities of the NPU.

Mu also adds three key transformer upgrades, namely pre- and post-LN, RoPE, and GQA.

pre- and post-LN (Double-layer normalization): Standardize before and after each sub-layer.

RoPE (Rotary position embedding): Embed relative positions into attention through complex-valued rotation.

GQA (Grouped query attention): Group queries, and each group shares a set of keys and values.

Mu is trained on Azure Machine Learning using NVIDIA A100 GPUs. First, it is pre-trained on hundreds of billions of high-quality educational tokens, then distills knowledge from the Phi model, and combines specific task data and LoRA fine-tuning.

Finally, although Mu is only a tiny model with hundreds of millions of parameters, only one-tenth the size of Phi-3.5-mini, its performance is comparable, and it can handle long input contexts and output quickly.

In addition, to operate efficiently on the device side, Mu uses the model quantization technology PTQ designed specifically for the NPU on Copilot+PC, converting model weights and activations from floating-point to integer representation, mainly 8-bit and 16-bit.

Microsoft also collaborates with chip partners for optimization, including adjusting mathematical operators and aligning with the execution mode of specific hardware, enabling efficient inference on edge devices.

For example, the following shows Mu running a Q&A task on an edge device using Wikipedia.

In the settings, Mu is fine-tuned by expanding the training samples to 3.6M and increasing the number of settings from 50 to hundreds, achieving a response time of less than 500 milliseconds and meeting the accuracy requirements.

In addition, this model is more suitable for multi-word queries. For fuzzy input of short words or partial words, the "Settings" app will continue to display lexical and semantic search results in the search box.

Frequent Recent Actions of Copilot+PC

Copilot+PC is Microsoft's ambitious project designed for the next wave of personal computing. It is powered by NPU and driven by AI, aiming to simplify and reshape the user's workflow, and has successively launched a variety of AI tools.

For example, the Recall function can help users trace back the current steps within seconds, quickly find and return to applications, websites, images, or documents.

Driven by AI, it also has better natural language search capabilities, regardless of whether the target document or image is in the File Explorer or the settings.

The recently launched Click to Do function can help quickly save text or images. You can copy text from an image to a summary text, or even quickly remove objects or backgrounds from an image, improving work efficiency while keeping the workflow running.

Now you can use Click to Do to quickly arrange a meeting or initiate a chat in Microsoft Teams with the recognized email while continuing to work, or be lazy and directly send the table information on the screen to Microsoft Excel. Your work efficiency will skyrocket.

In addition, Copilot+PC has recently introduced new features specifically for AI photo and drawing editing tools.

For example, Photos relight can relight photos. Just click to select your favorite built-in lighting preset to add creative effects to the photo.

A single light source can also be adjusted. Users can change the direction and proximity of the light towards the focus, as well as adjust color preferences and effect intensity.

Microsoft has added a sticker generator to the Paint application. A simple text prompt can create custom digital stickers, and you can also use AI to locate and edit a single specific element on the canvas.

After taking a screenshot, do you always need to crop it repeatedly to meet your needs?

Now Copilot+PC has launched an AI tool for perfect screenshots, which can automatically capture the screen content area and align the view to highlight the content.

In addition, there is a text extractor that can directly extract and copy text from an image, and a color picker that can capture the RGB value of a color from any position on the screen...

It can be said that Microsoft is now making drastic improvements to Copilot+PC, hoping that the user experience can be more intuitive, accessible, and useful.

However, just like the common voice of netizens, when can all Windows users enjoy these features?

So, will you buy a Copilot+PC for these new features?

Reference links:

[1]https://www.thurrott.com/windows/windows-11/322465/the-settings-agent-in-windows-11-has-its-own-ai-model

[2]https://blogs.windows.com/windowsexperience/2025/06/23/introducing-mu-language-model-and-how-it-enabled-the-agent-in-windows-settings/

[3]https://blogs.windows.com/windowsexperience/2025/05/06/introducing-a-new-generation-of-windows-experiences/

[4]https://blogs.windows.com/windows-insider/2025/06/13/announcing-windows-11-insider-preview-build-26200-5651-dev-channel/

This article is from the WeChat official account "Quantum Bit". The author is Lu Yu. 36Kr is published with authorization.