Phi-3

Available · Language · Milestone

Phi-3 is a family of instruction-tuned small language models from Microsoft in three sizes: Phi-3-mini (3.8 billion parameters), released on 23 April 2024, and Phi-3-small (7 billion) and Phi-3-medium (14 billion), released at Microsoft Build on 21 May 2024. Each size offers up to 128K tokens of context; weights are on Hugging Face, and mini and medium were also hosted in Azure. [1] [2] [3] Primary source

Timeline of Phi-3 →

Claims and evidence

  • Released Primary source[1]page date; Starting today, Phi-3-mini
  • Retired from Microsoft Foundry Primary source[4]Microsoft table
  • Status AvailableWeights of all three sizes are still downloadable from Microsoft's Hugging Face organisation (checked 2026-10-01). Retired from Microsoft Foundry on 2025-08-30 (see milestones). Primary source[3] [5] [6]
  • Successor of Phi-2The technical report calls phi-2 its previous model and says the Phi-3 training data is a scaled-up version of the phi-2 data. Primary source[7]abstract; Introduction
  • Change · Size Offered in three sizes, 3.8 billion (mini), 7 billion (small) and 14 billion (medium) parameters, against 2.7 billion for Phi-2.Compared with Phi-2 Primary source[2] [8]
  • Change · Context length Context window of up to 128K tokens, compared with 2,048 tokens for Phi-2; each size also has a short-context version.Compared with Phi-2 Primary source[1] [2] [8]
  • Change · Other Released as instruction-tuned models, post-trained with supervised fine-tuning and direct preference optimization; Phi-2 was a base model.Compared with Phi-2 Primary source[3] [9]
  • Change · Training data Trained on a scaled-up version of the phi-2 data, made of heavily filtered public web data and synthetic data; phi-3-mini saw 3.3 trillion tokens.Compared with Phi-2 Primary source[7]
  • Input text Primary source[3]Model, Inputs [5]Model, Inputs [6]Model, Inputs
  • Output text Primary source[3]Model, Outputs [5]Model, Outputs [6]Model, Outputs
  • Open weights Yes Primary source[3]License [5]License [6]License
  • Context window 128K tokensMaximum for every size; each size also has a short-context version (4K for mini and medium, 8K for small). Primary source[1] [2]The Phi-3 family
  • Access API, Open-weights downloadAt Build 2024 Phi-3-mini and Phi-3-medium were offered through Azure AI models as a service. Primary source[1] [2]Azure AI's models as a service offering

Lineage

Predecessors

  • Phi-2 · 12 December 2023

Successors

Based on

Not derived from another model.

Variants and derived

Siblings

None recorded.

All ancestors

All descendants

Variants

Phi-3-mini

  • Released Primary source[1]Starting today, Phi-3-mini
  • Variant Inline variant in this record.Offered in 4K and 128K context-length versions. Microsoft published a same-name update of the Phi-3-mini weights in June 2024 (model card release notes); it is not a separate record. Primary source[1] [3]

Also known as: Phi-3-mini-4k-instruct, Phi-3-mini-128k-instruct

Differs in:

  • Parameters: 3.8B Evidence not assessed [1] [3]Model Summary

Phi-3-small

  • Announced Primary source[1]In the coming weeks
  • Released Primary source[2] [5]Release dates
  • Variant Inline variant in this record.Offered in 8K and 128K context-length versions. Primary source[2] [5]

Also known as: Phi-3-small-8k-instruct, Phi-3-small-128k-instruct

Differs in:

  • Parameters: 7B Evidence not assessed [2]The Phi-3 family [5]Model Summary

Phi-3-medium

  • Announced Primary source[1]In the coming weeks
  • Released Primary source[2] [6]Release dates
  • Variant Inline variant in this record.Offered in 4K and 128K context-length versions. Primary source[2] [6]

Also known as: Phi-3-medium-4k-instruct, Phi-3-medium-128k-instruct

Differs in:

  • Parameters: 14B Evidence not assessed [2]The Phi-3 family [6]Model Summary

Related AI Radar coverage

AI Radar coverage starts in June 2026; no coverage linked yet.

All model releases from Microsoft on AI Radar →