News
ReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior ReleaseReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior Release
EveryLLM
imgPhi

Phi

Microsoft's small open-weights language models that run on a single device

Phi is a line of compact language models from Microsoft, first shown in June 2023. The weights are published so anyone can run them on their own hardware, including phones, and the newest member adds vision and selective reasoning.

Share on WhatsApp

  • First release, Phi-1, appeared on 21 June 2023 with 1.3 billion parameters.
  • Phi-3-mini packs 3.8 billion parameters and can run on a phone.
  • Weights are published under the MIT License.
  • Phi-4-reasoning-vision-15B, from March 2026, handles images and chooses when to reason.

What it is

Phi is Microsoft's family of small language models. Unlike large systems that only run in a data centre, these are built to be downloaded and executed locally, even on a phone. The series began with Phi-1 in June 2023 and has grown through Phi-3 in 2024 to a 15 billion parameter vision-and-reasoning release in 2026. Weights are published under the MIT License, so developers can take them and build on them freely. The aim is capable language understanding at a size that fits ordinary hardware.

How it came about

Microsoft put out the first member, Phi-1, on 21 June 2023 with 1.3 billion parameters. Phi-3-mini followed in April 2024 as an open-weights release sized for phones, and it was also made available through Azure. The newest entry, Phi-4-reasoning-vision-15B, arrived in March 2026. The New York Times read the Phi-3-mini launch as Microsoft moving to compete in the compact model space.

Inside the model

The line is defined by size rather than a single published architecture. Phi-1 carried 1.3 billion parameters and Phi-3-mini 3.8 billion, small enough to sit on a phone. The 2026 release, Phi-4-reasoning-vision-15B, holds 15 billion parameters and handles both text and images. Its distinguishing feature is a choice made at run time: it decides whether to show a chain of reasoning before answering or to reply straight away. Training data, layer counts and compute figures are not published here.

What it can do

Phi-3-mini was judged by WIRED and Ars Technica to reach roughly the level of GPT-3.5, a much larger model. Its compact footprint means it can be run on a phone rather than a server. Phi-4-reasoning-vision-15B adds image input and can pick when a problem deserves an explicit reasoning step. No benchmark tables are given for the family.

How to use it

All the models are open-weights under the MIT License, so the files can be downloaded and run locally. Phi-3-mini is additionally offered through Microsoft's Azure platform. The official page sits on the Azure site.

Reception and debate

The New York Times framed Phi-3-mini as Microsoft's bid for the small-model market. Two technology outlets, WIRED and Ars Technica, judged its output to be near GPT-3.5 in quality despite its much smaller size. No lawsuits, bans or safety incidents are recorded for the series.

Why it matters

Phi showed that a model small enough for a handset could hold its own against far bigger systems, and it made open weights the default for the line. The progression from a 1.3 billion parameter text model to a 15 billion parameter vision-and-reasoning one traces how quickly compact models gained new capabilities.

Questions people ask about Phi

What is Phi?

Phi is Microsoft's family of small language models. They are published as open weights under the MIT License and are designed to run on local hardware, including phones, rather than only in the cloud.

How big are the models?

The first, Phi-1, had 1.3 billion parameters. Phi-3-mini has 3.8 billion. The March 2026 release, Phi-4-reasoning-vision-15B, has 15 billion parameters and also accepts images.

How capable is Phi-3-mini?

WIRED and Ars Technica judged it to be close to GPT-3.5 in quality, even though it is far smaller and can be run on a phone.

How do I get the models?

The weights are downloadable under the MIT License. Phi-3-mini is also available through Microsoft's Azure platform.

What is special about the newest release?

Phi-4-reasoning-vision-15B, released in March 2026, takes images as well as text and decides at run time whether to lay out a reasoning process before answering or to skip it.

Good reading and sources

Where this page comes from

The facts on this page come from the Wikipedia article "Phi (language model)", written by its contributors and used under the CC BY-SA 4.0 licence, and from its Wikidata record (CC0). EveryLLM rewrote the text in its own words, with AI assistance, and compares every page with the article to make sure no passage is copied. It may be reused under the same CC BY-SA 4.0 licence with a link back.

It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.

The weekly digest

Every Monday: the model of the day, new pages on EveryLLM, and the conferences and anniversaries of the week ahead.