Microsoft has quietly introduced a powerful new AI feature in Windows 11 that could reshape how users interact with their PCs. Meet Mu, a compact, on-device small language model (SLM) designed to power AI agents within the Windows 11 Settings app. While it may not have the flash of Copilot or ChatGPT, Mu represents a significant leap in making everyday computing more intuitive, efficient, and private – especially for users in South Africa navigating multilingual environments and diverse device ecosystems.
Mu is a 330-million parameter encoder-decoder language model built specifically for Copilot+ PCs. Unlike cloud-based AI models, Mu runs entirely on-device using the Neural Processing Unit (NPU), enabling it to respond to natural language queries in under 500 milliseconds. Think of it as a smart assistant embedded directly into your system settings – ready to help you adjust brightness, toggle battery saver, or change accessibility options, all through simple voice or text commands.
South Africa’s tech landscape is marked by a growing demand for AI-driven productivity tools that are both accessible and secure. Mu’s on-device architecture means:
- No internet required: Ideal for users in areas with limited connectivity
- Enhanced privacy: Since data never leaves the device, it aligns with growing concerns around cloud-based data surveillance
- Multilingual potential: While currently English-focused, Mu’s architecture is designed to support additional languages in future updates – opening doors for isiZulu, Afrikaans, and other local languages
Mu is integrated directly into the Settings search box on supported Windows 11 builds. Users can type or speak commands like “reduce screen brightness at night” or “enable dark mode,” and Mu will either guide them to the correct setting or apply the change automatically (with permission).
Its encoder-decoder design allows it to process input and generate output more efficiently than traditional decoder-only models. This results in:
- Lower latency
- Higher throughput
- Reduced memory usage
These optimizations are crucial for real-time, on-device performance – especially on laptops with limited resources.
Mu is part of Microsoft’s broader push toward small, task-specific AI models that can run locally. It was trained using Microsoft’s Phi model family and fine-tuned with over 3.6 million examples to handle hundreds of system settings. The company plans to expand Mu’s capabilities to more devices, including those powered by Intel and AMD NPUs, and to support more languages and settings over time.
