TellWorldAI World Archive
Ask中文
News
Models
Organisations
People
Papers & concepts
Products & agents
Chips & robots
History
Topics
Ask中文

Home›Models

Model

Phi-3

Phi-3 is Microsoft's family of small language models, launched in April 2024 with Phi-3-mini, a 3.8-billion-parameter model small enough to run on a phone.

Status AvailableModality Text, ImageOpen weights YesContext (K tokens) 128 Era Generative AI for everyone

Parameters
14B
License
MIT
Access
API, Download
Release date
Developer
Microsoft

What it is

Phi-3 is a family of small language models from Microsoft. Its technical report was posted on 22 April 2024, and the first model, Phi-3-mini, launched on 23 April 2024.[1][2][5] Phi-3-mini has 3.8 billion parameters. The family later grew to include Phi-3-small, Phi-3-medium and Phi-3-vision, and these models are dense decoder-only Transformers, a common architecture for text generation.[1][4]

The family was positioned as small enough to run locally. The technical report's title describes a highly capable language model that runs on a phone, and Phi-3-mini takes about 1.8 GB of memory when quantised to 4 bits.[1]

Origins

Phi-3 is the successor of Phi-2. Microsoft scaled up the Phi-2 data approach for Phi-3-mini, training it on 3.3 trillion tokens of heavily filtered public web data and synthetic data.[1] A token is a chunk of text, often part of a word. Phi-3-small (7B) and Phi-3-medium (14B) were trained on 4.8 trillion tokens.[1]

Phi-3 models are aligned with supervised fine-tuning and direct preference optimisation, two post-training methods that teach a model to follow instructions and prefer answers people rate highly.[4]

Key features

Phi-3-mini came in 4K and 128K context-length variants. Microsoft said it was the first model in its class to support up to 128K tokens, which means it can take in a long document or a long conversation at once.[2]

On benchmarks, Phi-3-mini scores 69% on MMLU, a multi-subject knowledge and reasoning test, and 8.38 on MT-bench, a test of multi-turn chat quality. The report says these results are comparable to Mixtral 8x7B and GPT-3.5.[1] Phi-3-small scores 75% and Phi-3-medium 78% on MMLU.[1]

Microsoft said Phi-3-small beat GPT-3.5 Turbo and Phi-3-medium beat Gemini 1.0 Pro on its benchmarks. The company also noted that the models are weaker on factual-knowledge tests such as TriviaQA.[2][3]

Phi-3 was released with open weights under the MIT licence on Azure AI Studio, Hugging Face and Ollama.[2][4]

On 21 May 2024, at the Build conference, Microsoft added Phi-3-vision, a 4.2-billion-parameter multimodal model that handles images as well as text. It also made Phi-3-small (8K and 128K context) and Phi-3-medium (4K and 128K) available on Azure.[3]

A June 2024 update to Phi-3-mini raised its JSON-output score from 1.9 to 60.1 and its RULER long-context average from 68.8 to 84.6.[4]

Place in the family

Phi-3 sits before the Phi-3.5 update, which Microsoft released in August 2024.[1] The family combined a small text model, larger members and a multimodal variant, and it extended small models to longer contexts while keeping them able to run on everyday hardware.[1][2]

CapabilitiesLarge language modelsMultimodal

ApplicationsConsumer apps

Ecosystem & regionOpen source

Version history

7
  1. 2023
    • Jun 20, 2023 Phi-1

      Phi-1 is a 1.3-billion-parameter Transformer language model for Python from Microsoft Research, which showed that filtered and synthetic "textbook quality" training data can partly substitute for model and data scale.

    • Sep 11, 2023 Phi-1.5

      Phi-1.5 is a 1.3-billion-parameter language model from Microsoft Research, trained on curated synthetic "textbook" text rather than web crawls; its report claims natural-language performance comparable to models five times larger.

    • Dec 12, 2023 Phi-2

      Phi-2 is a 2.7-billion-parameter base language model from Microsoft Research, announced on 12 December 2023. Microsoft reported that it matched or beat models up to 25 times larger on complex benchmarks, making it a reference point for small-model training.

  2. 2024
    • Apr 23, 2024 Phi-3 · This page

      Phi-3 is Microsoft's family of small language models, launched in April 2024 with Phi-3-mini, a 3.8-billion-parameter model small enough to run on a phone.

    • Aug 2024 Phi-3.5

      Phi-3.5 is Microsoft's August 2024 update to its small Phi models, released in three versions — mini, vision and the line's first mixture-of-experts model — under the MIT licence.

    • Dec 12, 2024 Phi-4

      Phi-4 is Microsoft's 14-billion-parameter dense decoder-only language model, built with a data-quality-first recipe heavy on synthetic data and released with full weights under the MIT licence in January 2025.

  3. 2025
    • Feb 26, 2025 Phi-4-mini and Phi-4-multimodal

      Phi-4-mini is a 3.8B-parameter text model and Phi-4-multimodal a 5.6B model that adds images and speech, both released by Microsoft in February 2025 as the small, efficiency-focused members of the Phi-4 family.

20241

  1. ReleaseMicrosoft launches Phi-3-mini

    On 23 April 2024 Microsoft introduced the Phi-3 family of small language models, starting with Phi-3-mini, released on Azure AI Studio, Hugging Face and Ollama.

Developed by1

Microsoft

Successor of1

Phi-2

Succeeded by1

Phi-3.5

Uses1

Transformer

Notes and sources

  1. Phi-3 Technical Report: A Highly Capable Language Model Locally on Your Phonehttps://arxiv.org/abs/2404.14219
  2. Introducing Phi-3: Redefining what's possible with SLMshttps://azure.microsoft.com/en-us/blog/introducing-phi-3-redefining-whats-possible-with-slms/
  3. New models added to the Phi-3 family, available on Microsoft Azurehttps://azure.microsoft.com/en-us/blog/new-models-added-to-the-phi-3-family-available-on-microsoft-azure/
  4. microsoft/Phi-3-mini-128k-instruct model cardhttps://huggingface.co/microsoft/Phi-3-mini-128k-instruct
  5. Phi (language model)https://en.wikipedia.org/wiki/Phi_(language_model)

Cite this page: Phi-3. TellWorld, https://www.tellworld.net/models/phi-3. Updated 2026-11-08