Microsoft Unveils Mu AI Model for Local Device Use
Microsoft has introduced Mu, a new artificial intelligence model designed to run locally on devices. The model powers AI agents in Windows 11 settings. It enables users to describe actions. This allows AI to navigate or perform tasks autonomously.
Key Features
- On-Device Operation: Mu runs entirely on compatible Copilot+ PCs using the device’s neural processing unit (NPU).
- High Efficiency: Responds at over 100 tokens per second, meeting demanding user experience requirements.
- Small Language Model: Built on a transformer-based encoder-decoder architecture with 330 million token parameters.
Mu’s Development
- Training: Mu was trained using A100 GPUs on Azure Machine Learning and distilled from Microsoft’s Phi models.
- Optimisation: Fine-tuned via low-rank adaptation (LoRA) methods and task-specific data to improve efficiency.
- Performance: Performs similarly to Phi-3.5-mini despite being one-tenth the size.
Applications
- Windows Settings: Mu powers AI agents in Windows 11 settings. This allows users to describe actions. AI can now navigate or perform tasks autonomously.
- Future Development: Microsoft continues to refine Mu’s performance, addressing challenges like language-based gaps and vague queries.