Phi is Microsoft’s family of small language models, designed to deliver strong capability at compact sizes that can run on modest hardware. The Phi-3 Technical Report, “Phi-3 Technical Report: A Highly Capable Language Model Locally on Your Phone” (submitted April 22, 2024, by Microsoft researchers), introduces phi-3-mini, a 3.8 billion parameter model the authors report as comparable to much larger systems like Mixtral 8x7B and GPT-3.5 while small enough to run on a phone. The report also describes phi-3-small (7B) and phi-3-medium (14B) and a Phi-3.5 series including a multimodal vision variant.
The Phi approach emphasizes high-quality filtered and synthetic training data to get more capability per parameter. The line has since advanced to Phi-4: Microsoft’s own Phi-4 model card describes it as “a state-of-the-art open model” of 14 billion parameters in a dense decoder-only transformer, “built upon a blend of synthetic datasets, data from filtered public domain websites, and acquired academic books and Q&A datasets,” with a focus on advanced reasoning and math. Microsoft has extended the Phi-4 generation with smaller and specialized variants (such as Phi-4-mini, Phi-4-multimodal, and Phi-4 reasoning models); consult Microsoft’s current pages for the live lineup.
Distribution includes openly available model releases for local and on-device use.
Why business readers should care: Phi shows that carefully trained small models can rival far larger ones on many tasks, enabling cheaper, private, on-device AI that does not require sending data to a large hosted model.