Microsoft MAI models (series)

MAI is the in-house model family of Microsoft AI, the group led by Mustafa Suleyman. The family entry point lists each model in order, with its category, the availability its own announcement states, and the claims Microsoft makes. Benchmark figures are Microsoft’s; no independent evaluation is recorded here.

Lineage

DateReleaseNotes
2026-04-02MAI-Transcribe-1, MAI-Voice-1, MAI-Image-2First generation, launched in Microsoft Foundry ([[../../../RESOURCES/COMPANIES/microsoft
2026-06 (Build)MAI-Code-1-FlashCoding model for GitHub Copilot; the first MAI coding line
2026-09-04MAI Image 2.6Post “Pushing the quality-cost frontier with MAI Image 2.6”
2026-11-02MAI-Image-2.7Foundry catalogue version date
2026-11-17Ignite batch (below)Seven models announced; eighth item on the Ignite page is hardware

The Ignite 2026 batch (2026-11-17)

ModelCategoryStated availabilityMicrosoft’s headline claimNote
MAI Code 2CodingGitHub Copilot; Foundry not stated98% of GPT-6.1 Sol’s performance at 40% of its relative cost[[./mai-code-2
MAI Code 2 FlashCoding (high volume)GitHub Copilot; limited-access preview in FoundryComparable to GPT-6 Luna at 65% of its relative cost[[./mai-code-2-flash
MAI Cyber 2 FlashSecurity (agentic defence)Exclusive to Microsoft security productsBeats GPT-6 Astra on CyberGym-E2E (92.3% vs 91.5%)[[./mai-cyber-2-flash
MAI Image 2.7Image generation and editingFoundry, preview”#1 on Artificial Analysis image editing”[[../image-generation/mai-image-2-7
MAI Realtime 1Real-time speech-to-speechMAI Labs trial; Foundry limited-access previewOutperforms GPT-Realtime 2.1[[../audio/mai-realtime-1
MAI Transcribe 2Speech-to-textGenerally available in FoundryDragon Copilot Transcribe, built on it, halves multilingual error rates[[../audio/mai-transcribe-2
Decision 1.1Routing and decisionNot confirmed in the primary text50 times faster than GPT-6.1 Sol[[./decision-1-dot-1

The Ignite page counts seven models but lists eight items: the eighth, Maia 200, is an AI accelerator chip and is not part of the family.

Source limits

  • Every performance and cost figure is Microsoft’s own. Cost claims are relative percentages of competitors’ cost, not published prices.
  • Benchmark conditions differ between posts. Each model note records the conditions its own post states.
  • The announcement pages were read in summary form; the full text of some posts was not available to the review.
  • Secondary outlets give different model-name spellings and availability details. This series note follows the Microsoft pages.

Sources