device efficiency

Microsoft has introduced a small language model named Mu, designed to perform complex language tasks directly on devices like Copilot+ PCs, operating on a device's Neural Processing Unit. This model improves response times and reduces power and memory consumption compared to larger cloud-based AI models. Mu is developed using insights from Microsoft's Phi models and high-quality educational data, employing techniques such as distillation and low-rank adaptation for fine-tuning. Mu features an encoder–decoder architecture that separates input and output processes, enhancing speed and minimizing latency. It incorporates innovations like rotary positional embeddings, grouped-query attention, and dual LayerNorm. Microsoft also utilized quantization to reduce memory usage and improve speed without sacrificing accuracy, making Mu ideal for on-device applications. To enhance the AI agent in Windows Settings, Mu was fine-tuned with over 3.6 million examples, improving its ability to understand and manage system settings. The model addresses ambiguous commands and provides rapid assistance with precise multi-word commands, while maintaining a compact size that is one-tenth that of similar AI tools. Analysts on Wall Street have a positive outlook on MSFT stock, giving it a Strong Buy consensus rating based on 30 Buys and five Holds, with an average price target of 6.14 per share, indicating a potential upside of 6%.