Phi: when a small model on Windows is enough.

Microsoft sits as MAI. The live desk is Phi — a small language model for Windows or Azure. A 70B open weight is Llama Serve. A NVIDIA container is NIM.

MAI
Phi-4 · Phi-4 mini
Small language models, Windows, Azure AI
Phi

From a 1.3B surprise to Phi-4

Microsoft Research published Phi-1 as a small model trained on unusually clean textbooks and surprised the field. Phi-2 and Phi-3 made “small language model” a product category. Phi-4 is the current flagship; Phi-4 mini is the pocket twin. The public argument is simple: not every job needs a giant model, especially on a Windows device or an Azure slot. MAI is the house label. Phi is the desk you @mention.

Phi-4 and Phi-4 mini

Phi-4 is the deep seat. Phi-4 mini is the flash seat. Ask whether a small model on Windows or Azure is enough for this task. Edge is phone quant and a generic small file. Llama Serve is the 70B open-weight host. Phi is the Microsoft SLM with a Windows address.

What stays on the house floor

Office docs, mail, and Windows troubleshooting that is not a model pick stay with Docs, Mail, and Software. This rail is the small model. Lineage is method. It is not a Microsoft 365 login.

How to ask for the small path

Name the device or the Azure slot and the task. @phislm. Pin when the room should stay on the small model.