Mu is a 330 million parameter encoder-decoder language model launched by Microsoft, specially optimized for NPU and edge devices. It drives AI agents in Windows settings on Copilot+ PCs, converting natural language instructions into system actions in real time, achieving inference speeds of 200+ tokens/s on Surface Laptop 7.

Mu: Microsoft’s end-side small parameter language model specially created for NPU

Mu’s core parameters and statistics

Mu is a small parameter language model (SLM) officially released by Microsoft's Windows Applied Sciences team in June 2025. It focuses on running efficiently on NPU (Neural Processing Unit) and edge devices. It is built into the Windows settings of the Copilot+ PC and serves as the inference engine for the AI ​​agent, mapping the user's natural language instructions into system settings operations in real time.

Projects	Public Information
Developer	Microsoft (Microsoft)
Model Architecture	Encoder-Decoder Transformer
Parameter size	330 million (330M)
Inference speed	100+ tokens/s (NPU), 200+ tokens/s (Surface Laptop 7)
Input context	Tens of thousands of tokens
Output rate	More than 100 tokens/s
Deployment form	Device-side inference, completely offloaded to NPU
Quantization method	PTQ (post-training quantization), 8-bit / 16-bit integer
Training Hardware	A100 GPU (Azure Machine Learning)
Pre-training data	Hundreds of billions of high-quality education tokens
Distilled Source	Microsoft Phi Series Models
Fine-tuning method	LoRA (low-rank adaptation)
Supported Platforms	Windows 11 (Copilot+ PC, Dev Channel Insider Preview)
Publishing method	Integrated into Windows settings search box, no independent installation required

Positioning difference: Compared with Phi-3.5-mini (a small parameter model also produced by Microsoft), Mu parameter size is about 10 times smaller (330M vs 3.3B), but its performance on specific tasks (such as CodeXGlue code understanding) is close. It does not pursue general dialogue capabilities, but focuses on "input-output mapping" tasks - the most typical scenario is "understanding the user's system setting intentions and converting them into precise operating instructions."

Implications of architecture choice: Mu uses an encoder-decoder rather than a pure decoder architecture. The encoder processes all inputs at once to generate a potential representation, and the decoder only generates output based on this representation, avoiding the computational waste of reprocessing all inputs every time the decoder model generates a token. Official data shows that on the same level of NPU, the first token delay of the encoder-decoder architecture is reduced by about 47%, and the decoding speed is increased by about 4.7 times. This is critical for interactive scenarios that require sub-second response, such as setting up a search box.

Mu’s users and market recognition

Mu is not a product open to the public for download, but is delivered to users as part of Windows 11 system capabilities, so its "market acceptance" is reflected in Microsoft's internal adoption decisions and feedback from the Windows Insider community.

Microsoft Internal Verification: Mu passed Microsoft's strict internal double acceptance of accuracy and latency. In the Windows setting agent scenario, Mu's LoRA fine-tuned version achieved a comprehensive accuracy of 0.738 (F1) on hundreds of system settings, and the response time was controlled within 500ms, meeting the interaction design requirements of the UX team. In contrast, although Phi LoRA has higher accuracy (0.815), it was defeated because the model was too large to meet the latency target - this shows that Mu's positioning is not "the strongest", but "just right under limited resources".

Task level comparison with Phi-3.5-mini:

Evaluation task	After fine-tuning Mu (330M)	After fine-tuning Phi-3.5-mini (3.3B)
SQUAD (Reading Comprehension)	0.692	0.846
CodeXGlue (code understanding)	0.934	0.930
Windows Settings Agent	0.738	0.815

Mu even surpasses Phi-3.5-mini, which has 10 times larger parameters, by a slight advantage in the code understanding task (CodeXGlue), indicating that the encoder-decoder architecture has a natural parameter efficiency advantage in "input-output mapping" tasks.

Scale of coverage: Windows Settings Agent supports hundreds of system settings, covering high-frequency scenarios such as screen brightness, mouse pointer size, accessibility, network configuration, and personalization. Through methods such as synthetic data generation, automatic annotation, prompt word engineering, noise injection, and intelligent sampling, Microsoft expanded the training data from the original less than 50 settings to hundreds of items, with a sample size of 3.6 million.

Ecological Position: Mu represents an industry trend - sinking AI capabilities from the cloud to the device, converging from general-purpose large models to task-specific small models. It benchmarks Apple's On-Device Intelligence and Google's Tensor NPU routes, but has a narrower angle of entry (system settings proxy) and a deeper delivery method (embedded in the operating system search box).

Mu’s cost advantage

Mu's cost structure is fundamentally different from the traditional API model or cloud LLM: it does not have a separate pricing page and is not billed by token.

C-side/End User: Mu is completely free for Windows users using Copilot+ PC. It comes pre-installed on eligible devices as part of Windows, with no subscription fees or pay-as-you-go. The only up-front cost is the hardware—requiring a Copilot+ PC with an NPU like Qualcomm Snapdragon X, AMD Ryzen AI, or Intel Core Ultra.

Developer/Ecosystem: Mu currently does not have a public API or model weight download channel. Third-party developers cannot directly call Mu’s inference capabilities, nor can they embed the model in their own applications. If you want to use Mu's technical route in similar scenarios, you need to refer to the architectural ideas disclosed by Microsoft in the blog (encoder-decoder + NPU adaptation + PTQ quantization + LoRA fine-tuning), reproduce it yourself or wait for Microsoft to open the model weight.

Enterprise/IT Deployment: Mu’s enterprise value is reflected in reducing IT support costs and improving employee self-service efficiency. Traditionally, modifying system settings requires users to manually navigate menus or contact the IT department; Mu has shortened this process to "enter natural language → automate execution" and is expected to reduce the time spent on a single setting operation from an average of 30-60 seconds to 5-10 seconds. For businesses with thousands of Copilot+ PCs, the cumulative time savings are significant. However, this benefit is entirely dependent on hardware upgrades (purchasing Copilot+ PC), which is a hidden cost—enterprises need to complete equipment upgrades before they can enjoy the efficiency dividends brought by Mu.

Cost comparison (deduction):

Cost dimension	Cloud LLM solution (such as GPT-4o)	Mu client-side solution
Single inference cost	~$0.01-0.03/time (API call)	0 (hardware purchased)
Latency	500-2000ms (network + inference)	<500ms (full local)
Privacy	Data leaving the country requires compliance review	Data does not leave the device
Available offline	No	Yes
Hardware threshold	None (any internet-enabled device)	Requires Copilot+ PC (NPU)

Mu’s main functions

Mu's capabilities are exposed to end users through the Windows Settings Agent. The core is to convert natural language instructions into system operations.

• Natural Language Settings Search: Enter daily expressions such as "brighten the screen", "make the mouse pointer larger", "turn on night mode", "switch to dark theme", etc. in the Windows settings search box. After Mu understands the intention, it will jump directly to the corresponding settings page or perform an operation without having to remember the setting path. This is different from traditional keyword searches - traditional searches only return the settings page name, while Mu understands semantic variations, such as "make the screen less glaring" and can match brightness/night mode and other possibilities.

• HUNDREDS OF SETTINGS COVERED: From display, sound, network, and personalization to accessibility, privacy, and security, Mu’s fine-tuned version covers hundreds of modifiable system settings in Windows. Microsoft uses synthetic data generation and diversity sampling technology to ensure correct response to low-frequency settings.

• Sub-second end-side inference: Benefiting from the encoder-decoder architecture and NPU hardware acceleration, Mu's end-to-end response time is stable within 500ms, and the measured output speed on Surface Laptop 7 can reach 200+ tokens/s. Users see results almost immediately after typing—no need to wait for a round trip to the cloud.

• Fuzzy Query Intelligent Coverage: For queries that are too short or have unclear intentions (such as entering only the word "brightness"), Mu will automatically downgrade to traditional lexical and semantic search, and display a candidate list of relevant settings in the search box instead of forcibly giving imprecise operations. This hybrid mechanism of "AI-based, search-based" effectively avoids the risk of misoperation.

• Multi-language support: Mu's training data covers a variety of natural languages. Users can input instructions in English, Chinese, Japanese, German and other languages, and the model can accurately understand and perform corresponding operations. This is an important basic capability for multilingual users and enterprise global deployment.

Mu’s model and version evolution

Mu's version history is closely related to its delivery rhythm in Windows. The current traceable nodes are as follows.

Mainline release

Version	Date	Description
Early experimental version	~2025 Q1 (no official precise date yet)	Prototype verification based on Phi series distillation, used to evaluate the feasibility of the encoder-decoder architecture in NPU scenarios
Official release version (v1.0)	2025-06-23	Officially released through Windows Blog and integrated into the Copilot+ PC Settings Agent of Windows 11 Insider Preview Build 26200.5651 (Dev Channel)

Key Milestones

• May 2025: Microsoft first mentioned the concept of setting up agents in the "New Generation Windows Experience" preview, without revealing the underlying model details at the time.
• June 13, 2025: Windows 11 Insider Preview Build 26200.5651 (Dev Channel) is released, Copilot+ PC users get the ability to set up agents for the first time, and Mu starts to be available in grayscale for Insider users.
• June 23, 2025: Microsoft's official blog officially introduces the technical details of the Mu language model, including architecture selection, training process, quantization strategy, and fine-tuning methods for setting up agents.

Follow-up planning

Microsoft said in a blog post that it will continue to collect feedback through the Windows Insider program to optimize the experience of setting up the agent. From a technical perspective, subsequent iterations of Mu may advance in the following directions:

• Cover more system setting items (expanded from hundreds to thousands)
• Support compound operations (such as "dim the screen and turn on night mode" multiple commands at once)
• Improved recognition accuracy for ultra-short queries (1-2 words)
• Expanded to other system interaction scenarios beyond Windows settings

Currently Mu has not released Windows-independent model weights or APIs, nor has it made public model cards or Hugging Face repositories.

Mu’s technical advantages

Mu's technical advantage lies not in parameter scale, but in full-link design and engineering optimization that "completes specific tasks with the minimum computational cost."

Choice of encoder-decoder architecture: Mu uses an encoder-decoder rather than the current mainstream pure decoder architecture. In setting up the agent's usage mode, user input is typically a one-time complete natural language query (such as "turn the screen brightness to 50%") rather than a step-by-step conversation. The encoder encodes the input once and the decoder generates the output, avoiding the redundant calculation of "reprocessing the complete context every time a token is generated" in the decoder model. According to actual measurements on Qualcomm Hexagon NPU, this choice brings about a 47% reduction in first token latency and a 4.7x increase in decoding speed.

NPU hardware-aware optimization: Mu does not simply deploy the model to the NPU, but adjusts it from the architecture design stage to the parallel computing characteristics and memory limitations of the NPU. Specific measures include:

• Choose hidden layer dimensions and feedforward network widths that align with the NPU's preferred tensor dimensions to ensure peak efficiency for matrix operations on the hardware.
• The encoder and decoder adopt a layer distribution of approximately 2:1 (such as 32-layer encoder + 12-layer decoder) to maximize task adaptation capabilities under the premise that the total number of parameters is fixed.
• Use weight sharing (input embedding and output embedding share the same set of weights) to reduce the number of parameters and improve encoding and decoding vocabulary consistency.

Three Transformer Upgrades:

• Dual LayerNorm: Perform normalization before and after each sub-layer to ensure stable activation value scale, smoother training process, and faster convergence.
• Rotated Position Embedding (RoPE): Encodes relative position information through complex domain rotation operations, enabling the model to have the ability to extrapolate sequences beyond the training length while avoiding the length limitations of traditional absolute position encoding.
• Group Query Attention (GQA): Share key-value pairs between attention head groups, reducing the total amount of attention parameters and memory usage, while retaining the diversity of multi-head attention. This is especially important with the limited memory bandwidth of NPUs.

Training link from distillation to instruction fine-tuning:

1. Pre-training phase: Use hundreds of billions of high-quality education tokens for pre-training on A100 GPU, using warmup-stable-decay learning rate scheduling and Muon optimizer.
2. Knowledge distillation stage: Distill knowledge from the larger Microsoft Phi model, allowing Mu with 330M parameters to capture the language understanding capabilities of the larger model.
3. Task fine-tuning stage: Use the LoRA (low-rank adaptation) method to specifically fine-tune the agent scene. The training data is constructed through synthetic generation, automatic annotation, diverse rewriting, and noise injection, covering various semantic variants of hundreds of system settings.

PTQ quantization and hardware joint optimization: Post-training quantization (PTQ) converts model weights and activations from floating point numbers to 8-bit and 16-bit integer representation, significantly reducing memory usage and computing requirements. Microsoft has in-depth collaboration with its three chip partners, AMD, Intel, and Qualcomm, to ensure that the quantized operations are fully adapted and achieve the best performance on each NPU. Finally, an inference throughput of 200+ tokens/s was achieved on Surface Laptop 7.

How to use Mu

Mu does not have an independent entrance, and all its functions are exposed through the Windows Settings search box.

Prerequisites:

• A Copilot+ PC with NPU (Qualcomm Snapdragon X / AMD Ryzen AI / Intel Core Ultra)
• Joined the Dev Channel of the Windows 11 Insider program and installed Build 26200.5651 or newer

Operating steps:

1. Click the "Start" button on the taskbar or press Win + I to open "Settings".
2. Enter a natural language command (such as "dim screen", "increase text size", "turn on Bluetooth") in the search box at the top of the settings window.
3. Mu analyzes the intent in real time on the client side - if the instruction is clear and involves automatically executable settings, the settings page automatically jumps to the corresponding item and highlights the operation area; if the instruction is vague or involves multiple settings, the search box displays a candidate list for the user to choose from.

Developer Access Method: Mu currently does not provide a public API or SDK. Developers who want to integrate similar capabilities into their own applications can refer to the technical solution disclosed on Microsoft's blog (encoder-decoder architecture + LoRA fine-tuning + PTQ quantization) and build their own inference pipeline based on the Phi series or equivalent small models. Microsoft has not announced whether it will release Mu to Hugging Face or Azure AI Foundry in the same way as the Phi series.

Mu’s Product Pricing

As an AI capability built into the Windows operating system, Mu does not have independent pricing.

• End User Cost: Free. Features come pre-installed on eligible Copilot+ PCs and require no separate purchase or subscription.
• Enterprise Cost: Implicit. Enterprises need to complete the hardware upgrade of Copilot+ PCs (usually purchased on a per-device basis, ranging from several thousand to ten thousand yuan per unit) before employees can use this feature. Existing non-Copilot+ PC devices cannot gain Mu capabilities via software update.
• Developer/Third Party Cost: Currently cannot be purchased or accessed directly. If you want to use similar capabilities in your own products, you need to develop it yourself or wait for Microsoft to open up the model weight.

In comparison, the cloud LLM solution has a clear pay-as-you-go cost in the API call scenario, but Mu’s advantage lies in zero marginal inference cost and privacy guarantee that the data does not leave the device. For IT managers, the decision to enable Mu or not is rarely a stand-alone decision - it depends on whether the business has already upgraded or plans to upgrade to Copilot+ PCs.

Mu application scenarios

Windows Settings Agent: This is the only scenario where Mu is currently officially implemented. Users can quickly adjust system settings through natural language in daily use, which is especially suitable for the following sub-scenarios:

• Users with poor vision or mobility issues can adjust accessibility features (magnifying glass, high contrast, narrator) through voice/text commands.
• New users unfamiliar with Windows Settings menus can quickly find and modify target settings using natural language.
• Multi-monitor users quickly specify brightness/resolution changes for a certain screen.
• IT support staff can quickly locate configuration items through precise instructions when assisting remotely.

Device-side real-time interaction: Mu's architectural design makes it naturally suitable for device-side inference tasks that require low latency and high throughput. In addition to setting up the agent, it can theoretically be extended to:

• Windows search enhancement: Integrate file search, application launch and system control into a unified natural language entrance.
• Quick operation suggestions: Learn the user's usage patterns and proactively recommend setting changes (such as "It is detected that you often project in the conference room. Do you want to switch to extended mode?").
• Accessibility assistance: Combined with voice input, it provides complete voice control system setup capabilities for users with motor disabilities.

Edge AI Prototype Reference: For hardware manufacturers and AI developers, Mu's public technical report can be used as a reference template for end-side model design - encoder-decoder is still an effective choice for processing "input-output mapping" type tasks on computing-constrained devices, and the combination of LoRA + PTQ can complete the adaptation of new tasks within a week.

Not suitable for scenarios:

• Mu is not suitable for conversational interaction, long text generation, creative writing or complex reasoning tasks - its architecture and training data are designed for "single input → precise output".
• Not suitable for non-Windows platforms: Currently Mu is completely bound to the Windows ecosystem and cannot run on Linux, macOS or mobile devices.
• Not suitable for scenarios that require customized model behavior: users cannot modify Mu’s inference parameters (temperature, top_p, etc.), nor can they switch to other fine-tuned versions.

Mu’s applicable groups

• Copilot+ PC owners: If you are a Windows 11 Insider and have a Copilot+ PC, Mu is a zero-cost, out-of-the-box system feature. It is suitable for users who often adjust system settings but cannot remember menu paths, as well as multi-tasking users who want to improve efficiency through voice/quick text operations.

• IT Administrators and Equipment Purchasing Decision Makers: Mu is an incremental value point for enterprise IT-enabled automation. If a team is evaluating the purchase of a Copilot+ PC, "reduced IT tickets due to self-service setup" can be factored into the ROI calculation. However, it should be noted that the premise of this benefit is that employees are willing to change their operating habits from "calling IT" to "searching themselves."

• Device-side AI developers and researchers: Mu’s public technical information (architecture selection, NPU adaptation strategy, quantification and distillation process) is an important reference for end-side model engineering. However, Mu itself currently does not have open model weights or APIs, and the directly reusable assets are limited. It is recommended to pay attention to whether Microsoft will open up models with similar capabilities through the Phi series or Azure AI Foundry.

• Not applicable to people:

  • Developers who need cross-platform end-side model capabilities (Mu is bound to Windows/NPU).
  • Users seeking general conversation or content generation capabilities (should choose general small models such as Phi, Gemma or Llama).
  • Teams looking for out-of-the-box API services (Mu has no public API and is only available through the Windows settings interface).
  • Users who need to use on-device AI on non-Copilot+ PCs (Mu requires NPU hardware support).

Summary and Outlook

Mu is a specific step in Microsoft's strategy of "sinking AI to the system level". It uses a combination of 330 million parameters, encoder-decoder architecture, and NPU full offloading technology to achieve sub-second natural language setting interaction on computing-constrained devices. Compared with general-purpose large models, Mu does not pursue broad dialogue capabilities, but "is adequate and extremely fast for narrow tasks" - this "less is more" design philosophy represents a pragmatic path for device-side AI.

Current Limitations:

• Narrow scope of functionality: only covers Windows settings and cannot handle more general system interactions such as file management, app launch, scheduling, etc.
• Hardware binding: Must rely on the NPU of Copilot+ PC, and existing PC users cannot obtain a similar experience through software updates.
• Ecological closure: model weights and APIs are not open, and third parties cannot reuse or expand.
• Still in the Insider stage: Features have not yet been pushed to the Windows 11 General Availability (GA) channel, and production availability is yet to be confirmed.

Key Observation Points:

• Whether Microsoft will expand Mu to other system interaction scenarios beyond settings (such as file search, control panel) in subsequent Windows versions.
• Whether Microsoft will open up Mu's model weights through the Phi series or Hugging Face so that developers can reproduce or fine-tune their own device-side models.
• Will the performance improvements of chip partners (AMD, Intel, Qualcomm) on next-generation NPUs allow Mu's response speed to further break through.
• The corresponding actions of Apple and Google in device-side AI (Apple Intelligence, Google Tensor) and the differentiated competition direction of Mu route.

Procurement/Adoption Risk Assessment: For individual users, as long as the hardware meets the conditions, they can experience it at zero cost and the risk is extremely low. For enterprises, the deployment cost is mainly reflected in Copilot+ PC hardware upgrades. It is recommended to gradually promote it when the device update cycle comes naturally, rather than upgrading individual functions in advance. It is recommended that this feature be pushed to the official version of Windows 11 before being fully promoted to avoid instability in the Insider version. For developers, it is not recommended to directly invest in product development based on Mu's current technical report - before Microsoft clears its open strategy, the Phi series is a more mature and accessible device-side model choice.

Related tools: , LangChain

Version Info

• Mu official version ：Officially released via Windows Blog and integrated into the Copilot+ PC Settings agent in Windows 11 Insider Preview Build 26200.5651 (Dev Channel). （2025-06-23）
• Early development version ：There is no official precise date yet. An early experimental version based on Phi series distillation and NPU adaptation, used to verify the feasibility of the encoder-decoder architecture in end-side scenarios. （~2025-Q1）