MAI-Thinking-1

Preview · Language, Reasoning · Milestone

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 a sparse mixture-of-experts model with 35 billion active and about 1 trillion total parameters, which Microsoft says it pre-trained from scratch without distillation from third-party models. [1] [2] [3] [4] Primary source

Timeline of MAI-Thinking-1 →

Claims and evidence

  • Announced Announced as one of seven MAI models; at announcement it was offered only as a private preview in Microsoft Foundry. Primary source[1]Our Models [2]dateline; Availability and access
  • Released Public preview in Microsoft Foundry. Before that, access was a private preview on request from 2026-06-02 (see milestones). Primary source[3]dateline; Updated as of August 12, 2026
  • Private preview in Microsoft Foundry The June version of the announcement says it is available in private preview on Microsoft Foundry that day; the model card describes early access through an interest form. Primary source[2]Availability and access [5]Distribution channels
  • Status PreviewPublic preview in Microsoft Foundry since 2026-08-12; no general availability found up to 2026-10-01. Primary source[6]heading; MAI-Thinking-1 at a glance [7]Microsoft models sold by Azure: MAI-Thinking-1 [3]Availability and access
      • Input text Primary source[5]Technical specs: Input formats [7]MAI-Thinking-1: Capabilities
      • Output textThe model card leaves output formats unspecified; the Foundry models page lists text output of up to 64,000 tokens. Primary source[7]MAI-Thinking-1: Capabilities
      • Feature Reasoning modeMicrosoft AI calls it a reasoning model; the API can return an encrypted reasoning state. Primary source[3]first paragraphs [6]introduction; Preserve reasoning state across turns
      • Feature Function calling Primary source[3]Enterprise ready [6]Use function calling and tools
      • Context window 256,000 tokens tokensThe how-to page, model card and technical report write 256K and the announcement 256k; output is capped at 64K tokens. Primary source[7]MAI-Thinking-1: Context length [6]Token limits and context window [3]Enterprise ready [4]1 Introduction
      • Parameters c. 35B-active, ~1T-totalSparse mixture-of-experts. The value is the total; the technical report writes 35B active / 1T total without the tilde. Primary source[3]Medium-sized model, with strong software engineering performance [4]Abstract; 2 Pre-training
      • Access APIThrough Microsoft Foundry (GlobalStandard deployment, chat completions API). Primary source[3]Availability and access [6]Prerequisites

      Lineage

      Predecessors

      No known predecessor.

      Successors

      No known successor.

      Based on

      Not derived from another model.

      Variants and derived

      Siblings

      None recorded.

      All ancestors

      None.

      All descendants

      Variants

      No variants recorded in this record.

      Related AI Radar coverage

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

      All model releases from Microsoft on AI Radar →