AI Radar

Microsoft models

Microsoft, founded in 1975 by Bill Gates and Paul Allen and based in Redmond, Washington, describes itself as a technology company that builds AI-powered platforms and tools. Microsoft Research developed the Phi small language models, released as open models under the MIT License, and the Microsoft AI organization, formed in 2024, develops the MAI model family. Known as Micro-Soft.

Chronology

  1. 2026

    4 models
    1. Released
      MAI-Thinking · Language, Reasoning · Milestone MAI-Thinking-1

      MAI-Thinking-1 is a reasoning model that Microsoft AI calls its flagship, announced on 2 June 2026 as a private preview and in public preview on the Microsoft Foundry API since 12 August 2026. It is…

      Primary source Preview · no news linked
    2. Released
      MAI-Code · Language, Multimodal, Code MAI-Code-1.1-Flash

      MAI-Code-1.1-Flash is a coding model from Microsoft AI, released on 11 August 2026 as the successor to MAI-Code-1-Flash and offered inside GitHub Copilot. It is a sparse mixture-of-experts model…

      • Adds native image input; MAI-Code-1-Flash accepted text only.
      • Microsoft says it uses about 25% fewer tokens per task and streams tokens about 25% faster in GitHub Copilot.
      Primary source Available · no news linked
    3. Released
      MAI-Code · Language, Code MAI-Code-1-Flash

      MAI-Code-1-Flash is a coding model from Microsoft AI, released on 2 June 2026 and offered only inside GitHub Copilot, where GitHub deprecated it across all Copilot experiences on 10 September 2026…

      • Smaller: 5B active and 137B total parameters, against 35B active and about 1T total for MAI-Thinking-1, from whose mid-training checkpoint it was trained.
      • Specialised for coding: trained further on synthetic agentic tasks and with reinforcement learning and the GitHub Copilot production harness; offered only inside GitHub Copilot.
      Primary source Retired · no news linked
    4. Released
      Phi reasoning · Language, Multimodal, Reasoning Phi-4-reasoning-vision-15B

      Phi-4-reasoning-vision-15B is a 15-billion-parameter open-weight multimodal reasoning model from Microsoft Research, released on 4 March 2026 under the MIT licence through Hugging Face, GitHub and…

      • Adds image input: a SigLIP-2 vision encoder turns images into tokens that are projected into the Phi-4-reasoning language backbone (mid-fusion).
      • Trained on a mix of reasoning and non-reasoning data, so one model can reason step by step or answer directly, selected with think and nothink mode tokens.
      Primary source Available · no news linked

  2. 2025

    8 models
    1. Released
      MAI · Language · Milestone MAI-1-preview

      MAI-1-preview is a mixture-of-experts language model that Microsoft AI calls its first foundation model trained end to end in-house, announced on 28 August 2025. It was offered for public testing on…

      Primary source Status unknown · no news linked Some details lack sources
    2. Released
      Phi reasoning · Language, Reasoning Phi-4-mini-flash-reasoning

      Phi-4-mini-flash-reasoning is a 3.8-billion-parameter open model from Microsoft for math reasoning with a 64K-token context, released on 9 July 2025 on Azure AI Foundry, the NVIDIA API Catalog and…

      • Replaces the dense Transformer with the SambaY decoder-hybrid-decoder architecture, which uses Gated Memory Units.
      • Microsoft says it decodes with higher throughput and lower latency than Phi-4-mini-reasoning, especially on long generations.
      Primary source Available · no news linked
    3. Released
      Phi · Language Phi-mini-MoE

      Phi-mini-MoE is an open-weights mixture-of-experts language model from Microsoft with 7.6 billion total and 2.4 billion active parameters, compressed from Phi-3.5-MoE with the SlimMoE method and…

      • Compressed to 7.6B total and 2.4B active parameters, from 41.9B total and 6.6B active in Phi-3.5-MoE.
      • Experts slimmed with the SlimMoE compression framework, with knowledge transferred through staged distillation.
      Primary source Available · no news linked
    4. Released
      Phi reasoning · Language, Reasoning Phi-4-mini-reasoning

      Phi-4-mini-reasoning is a 3.8-billion-parameter open-weight model from Microsoft for step-by-step math reasoning, released on 30 April 2025 on Azure AI Foundry and Hugging Face under the MIT…

      • Trained with distilled long chain-of-thought data, rollout DPO and reinforcement learning with verifiable reward to solve math problems step by step.
      • Lists English as its only supported language, where the Phi-4-mini model card lists 23.
      Primary source Available · no news linked
    5. Released
      Phi reasoning · Language, Reasoning Phi-4-reasoning

      Phi-4-reasoning is a 14-billion-parameter open-weight reasoning model from Microsoft, created by supervised fine-tuning of Phi-4 and released on 30 April 2025 on Azure AI Foundry and Hugging Face…

      • Fine-tuned to write out a chain of thought before giving a summarized answer.
      • Maximum length raised from 16K to 32K tokens to fit longer reasoning traces.
      Primary source Available · no news linked
    6. Released
      MAI · Language, Reasoning MAI-DS-R1

      MAI-DS-R1 is an open-weights reasoning model released by Microsoft on 17 April 2025: a version of DeepSeek-R1 that the Microsoft AI team post-trained to answer topics the original blocked and to…

      Primary source Available · no news linked
    7. Released
      Phi · Language Phi-4-mini

      Phi-4-mini is a 3.8-billion-parameter text model in Microsoft's Phi family of small language models, released on 26 February 2025 with Phi-4-multimodal. It was offered in Azure AI Foundry, the…

      • Vocabulary grows to 200K tokens, which Microsoft says is meant to serve multilingual applications better than Phi-3.5-mini.
      • Adds grouped-query attention and shared input and output embeddings compared with Phi-3.5-mini.
      Primary source Available · no news linked
    8. Released
      Phi multimodal · Language, Multimodal Phi-4-multimodal

      Phi-4-multimodal is a 5.6-billion-parameter model in Microsoft's Phi family that accepts text, image and audio input and produces text, released on 26 February 2025 with Phi-4-mini. It handles all…

      • Adds image and audio input to the Phi-4-mini language model, which stays frozen, through vision and speech encoders with LoRA adapters.
      • Adds speech and audio input; with the Phi-3 series, users had to put a separate speech recognition model in front of the Vision model.
      Primary source Available · no news linked

  3. 2024

    5 models
    1. Released
      Phi · Language · Milestone Phi-4

      Phi-4 is a 14 billion parameter small language model from Microsoft, released on 12 December 2024 and aimed at complex reasoning such as math as well as conventional language tasks. Microsoft offers…

      • Training data extends the Phi-3 data with new synthetic textbook-like data and acquired academic books and Q&A datasets; synthetic data is used throughout.
      • Context length is 16K tokens, compared with 128K tokens for the Phi-3.5 models.
      Primary source Available · no news linked
    2. Day not given
      Phi · Language Phi-3.5

      Phi-3.5 is a Microsoft release of two open-weight text models announced on 22 August 2024: Phi-3.5-mini (3.8 billion parameters) and the mixture-of-experts Phi-3.5-MoE (42 billion in total, 6.6…

      • Adds multilingual support: the Phi-3.5-mini card lists 23 languages, and Microsoft says Phi-3.5-MoE supports more than 20.
      • Adds a mixture-of-experts model, Phi-3.5-MoE, with 16 experts of 3.8 billion parameters each; 6.6 billion of its 42 billion parameters are active.
      Primary source Available · no news linked
    3. Day not given
      Phi multimodal · Language, Multimodal Phi-3.5-vision

      Phi-3.5-vision is a 4.2 billion parameter multimodal model from Microsoft that takes text and images and returns text, announced on 22 August 2024 with Phi-3.5-mini and Phi-3.5-MoE. It adds…

      • Adds multi-frame input: the model can compare several images and summarise multi-image sets and video clips.
      Primary source Available · no news linked
    4. Released
      Phi multimodal · Language, Multimodal Phi-3-vision

      Phi-3-vision is a 4.2 billion parameter multimodal model from Microsoft that accepts text and images and returns text, the first multimodal model in the Phi-3 family. Introduced at Microsoft Build…

      • Adds image input to the Phi-3-mini language model through an image encoder, connector and projector, keeping a 128K-token context.
      Primary source Available · no news linked
    5. Released
      Phi · Language · Milestone Phi-3

      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…

      • Offered in three sizes, 3.8 billion (mini), 7 billion (small) and 14 billion (medium) parameters, against 2.7 billion for Phi-2.
      • Context window of up to 128K tokens, compared with 2,048 tokens for Phi-2; each size also has a short-context version.
      Primary source Available · no news linked

  4. 2023

    3 models
    1. Released
      Phi · Language · Milestone Phi-2

      Phi-2 is a 2.7 billion parameter base language model from Microsoft, announced by Satya Nadella at Microsoft Ignite in November 2023 as a scaled-up version of Phi-1.5. Microsoft Research released it…

      • Grows from 1.3 billion parameters in Phi-1.5 to 2.7 billion.
      • Adds new synthetic NLP texts and websites filtered for safety and educational value to the Phi-1.5 data sources.
      Primary source Available · no news linked
    2. Announced, release date unknown
      Phi · Language Phi-1.5

      Phi-1.5 is a 1.3 billion parameter base language model from Microsoft, presented in a September 2023 technical report that applies the phi-1 textbook-data approach to common-sense reasoning in…

      • Adds a new source of synthetic NLP texts to the phi-1 data sources, moving the focus from Python coding to common-sense reasoning in natural language.
      • The model card lists question-answer, chat and code prompt formats, whereas phi-1 was specialised for basic Python coding.
      Primary source Available · no news linked
    3. Announced, release date unknown
      Phi · Language, Code · Milestone Phi-1

      Phi-1 is a 1.3 billion parameter Transformer language model from Microsoft, specialised for basic Python coding and introduced in the paper Textbooks Are All You Need in June 2023. It was trained on…

      Primary source Available · no news linked

Lineage

  1. Phi-1
    Successors
    Phi-1.5
  2. Phi-1.5
    Predecessors
    Phi-1
    Successors
    Phi-2
  3. Phi-2
    Predecessors
    Phi-1.5
    Successors
    Phi-3
  4. Phi-3
    Predecessors
    Phi-2
    Successors
    Phi-3.5
    Variants and derived
    Phi-3-vision · derived: other
  5. Phi-3-vision
    Successors
    Phi-3.5-vision
    Based on
    Phi-3 · derived: other
  6. Phi-3.5
    Predecessors
    Phi-3
    Successors
    Phi-4; Phi-4-mini
    Variants and derived
    Phi-mini-MoE · derived: distillation
  7. Phi-3.5-vision
    Predecessors
    Phi-3-vision
  8. Phi-4
    Predecessors
    Phi-3.5
    Variants and derived
    Phi-4-reasoning · derived: fine tune
  9. Phi-4-mini
    Predecessors
    Phi-3.5
    Variants and derived
    Phi-4-mini-reasoning · derived: fine tune; Phi-4-multimodal · derived: adapter
  10. Phi-4-multimodal
    Predecessors
    Phi-3.5-vision
    Based on
    Phi-4-mini · derived: adapter
  11. MAI-DS-R1
    Based on
    DeepSeek-R1 (DeepSeek) · derived: fine tune
  12. Phi-4-mini-reasoning
    Based on
    Phi-4-mini · derived: fine tune; DeepSeek-R1 (DeepSeek) · derived: distillation
  13. Phi-4-reasoning
    Based on
    Phi-4 · derived: fine tune; o3-mini (OpenAI) · derived: distillation
    Variants and derived
    Phi-4-reasoning-vision-15B · derived: other
  14. Phi-mini-MoE
    Based on
    Phi-3.5 · derived: distillation
  15. Phi-4-mini-flash-reasoning
    Based on
    DeepSeek-R1 (DeepSeek) · derived: distillation
  16. MAI-1-preview

    No recorded relations.

  17. Phi-4-reasoning-vision-15B
    Based on
    Phi-4-reasoning · derived: other
  18. MAI-Code-1-Flash
    Based on
    MAI-Thinking-1 · derived: other
  19. MAI-Code-1.1-Flash
    Predecessors
    MAI-Code-1-Flash
    Based on
    MAI-Thinking-1 · derived: other
  20. MAI-Thinking-1
    Variants and derived
    MAI-Code-1.1-Flash · derived: other; MAI-Code-1-Flash · derived: other

Families

About the organisation

  • Description Microsoft, founded in 1975 by Bill Gates and Paul Allen and based in Redmond, Washington, describes itself as a technology company that builds AI-powered platforms and tools. Microsoft Research developed the Phi small language models, released as open models under the MIT License, and the Microsoft AI organization, formed in 2024, develops the MAI model family. Primary source[2] [1] [3] [4] [5] [6]
  • Former name Micro-Soft (from ) Primary source[1]1975: Microsoft is born

Company · https://www.microsoft.com · 20 models listed here

Sources

  1. [1] About Microsoft

    Other · Microsoft

    Provenance
    Primary
    Availability
    Active
    Last checked
    Original ↗ https://news.microsoft.com/about/ Living company history timeline on Microsoft Source. The 1975 entry says Gates and Allen started Microsoft, sometimes written Micro-Soft, to develop software for the Altair 8800. The HTML title reads About Microsoft - Stories.

  2. [2] Facts About Microsoft

    Other · Microsoft

    Provenance
    Primary
    Availability
    Active
    Last checked
    Original ↗ https://news.microsoft.com/facts-about-microsoft/ Living company fact sheet on Microsoft Source (corporate address, important dates, mission); content as seen on 2026-10-01. The HTML title reads Facts About Microsoft - Stories.

  3. [3] Phi-2: The surprising power of small language models

    Blog post · Microsoft · published

    Provenance
    Primary
    Availability
    Active
    Last checked
    Original ↗ https://www.microsoft.com/en-us/research/blog/phi-2-the-surprising-power-of-small-language-models/ Microsoft Research Blog; the page title ends with - Microsoft Research. A figure caption says Satya Nadella announced Phi-2 at Microsoft Ignite 2023.

  4. [4] Phi Open Models - Small Language Models

    Other · Microsoft

    Provenance
    Primary
    Availability
    Active
    Last checked
    Original ↗ https://azure.microsoft.com/en-us/products/phi Azure product page; the heading reads Phi open model family. The FAQ lists the Phi series from Phi-1 to Phi-4-multimodal and says Phi models are open source under the MIT License. The HTML title ends with the suffix | Microsoft Azure.

  5. [5] Mustafa Suleyman, DeepMind and Inflection Co-founder, joins Microsoft to lead Copilot

    Announcement · Microsoft · published

    Provenance
    Primary
    Availability
    Active
    Last checked
    Original ↗ https://blogs.microsoft.com/blog/2024/03/19/mustafa-suleyman-deepmind-and-inflection-co-founder-joins-microsoft-to-lead-copilot/ Official Microsoft Blog; Satya Nadella's message to employees announcing the new Microsoft AI organization.

  6. [6] Building a hill-climbing machine: Launching seven new MAI models

    Announcement · Microsoft · published

    Provenance
    Primary
    Availability
    Active
    Last checked
    Original ↗ https://microsoft.ai/news/building-a-hillclimbing-machine-launching-seven-new-mai-models/ Microsoft AI post by Mustafa Suleyman with the line Updated as of June 8, 2026 (page metadata modified 2026-07-29). It announces MAI-Thinking-1, MAI-Code-1-Flash and image, voice and transcription models, and calls them the MAI model family. HTML title ends with the suffix | Microsoft AI.