Might be it’s the shortest title of any Daily Post post, mainly referring to the one-word titles. That apart you would have guessed it right, it is yet another offering in the field of AI. The number of AI offerings since the ChatGPT hit the public domain has been humongous and way beyond what any technology even in humans has managed so far, and that too in such a short period of time. Additionally, all have been impact making in their own way, marking different milestones in the onward march of AI becoming more and more pervasive by the day.
Mu is another AI milestone and comes from Microsoft fold. This is in some way a continuation of Microsoft’s Copilot +PC featuring a Neural Processing Unit (NPU) capable of 40 TOPS or more. These devices came with Phi-Silica, an on-device SLM. This was able to deliver language intelligence capabilities to Microsoft apps. On 23rd of June, 2025, Microsoft went one step forward, when it revealed Mu, an on-device small language model built into Windows 11. What is the end goal of this exercise? It was to create an AI powered agent within the Settings app.
Taking it further, what is the Agent technically competent to do? It will understand users’ natural language queries and integrate it into the existing search box for a smooth user experience. Mu was made available to Windows Insiders recently in the Dev Channel with Copilot+PC. It can operate efficiently on NPU, delivering over 100 tokens per second, running locally. In a recent official blog post Microsoft explained how they designed and trained the Mu model. Technically, it is an encoder-decoder model.
First token latency was brought down by 47% because of the encoder-decoder approach. It has also been able to achieve 4.7x higher decoding speed compared to a decoder only model of similar size. Weight sharing in certain components reduces the total parameter count. Mu has been trained on Azure Machine Learning. What is most fascinating is that Mu is one-tenth the size of Phi-3.5-mini and is nearly comparable in performance, once similarly fine-tuned. The model is better suited for multi-word queries. AI is slowly becoming all pervasive in more ways than one.
EMBEDDED AI IS THE FUTURE.
Sanjay Sahay
Have a nice evening.