ai-trend-notifier
← wiki

$ cat wiki/models/mai-thinking-1.md

MAI-Thinking-1

modelupdated 2026-07-21created 2026-06-02

Spec

AttributeValue
DeveloperMicrosoft (MAI team, led by Mustafa Suleyman)
Releasednot yet (exact API/GA date not announced as of 2026-06-02)
Announced2026-06-02 (Microsoft Build 2026 keynote)
Context window128K tokens
Pricingunknown
LicenseProprietary
AvailabilityAzure AI Foundry, Copilot model picker
ArchitectureReasoning-specialized (reasoning model)
Active Parameters35B
Training ApproachTrained from scratch — no distillation from third-party frontier models

Release Date

2026-06-02 (announced at Build 2026). Exact API/GA date not announced.

What Is It

A reasoning-specialized model developed in-house by Microsoft's MAI (Microsoft AI) team. One of Microsoft's family of frontier models announced alongside MAI-Code-1/Polaris.

Key features:

  • Independent training: trained from scratch with no distillation from third-party models such as OpenAI GPT, Anthropic Claude, or Google Gemini → securing independence from competitors' IP
  • Reasoning-specialized: CoT/ToT-based multi-step planning — handling complex reasoning tasks
  • 35B active params: positioned as a mid-size model within the MAI family

Significance

  • The first in-house reasoning model Microsoft has released since lifting OpenAI partnership constraints (2026-04)
  • An independent entry into the reasoning model market at the tier of GPT-4o mini, Claude Haiku, and Gemini Flash
  • Expected to serve as the orchestration brain of agent platforms (Azure Agent Mesh, Windows Agent Framework)

Benchmarks

Not released (as of 2026-06-02). No specific benchmark numbers disclosed at the Build 2026 keynote.

Use Cases

  • Multi-step reasoning agent tasks
  • Task orchestration within Azure Agent Mesh
  • Planning for complex enterprise workflows

Compared To

ModelCharacteristics
Claude Opus 4.8Anthropic's strongest model, SWE-bench Pro 69.2%, larger than 35B
GPT-o3/o4 (OpenAI)Comparable reasoning model, parameter scale undisclosed
Gemini 3.1 Deep ThinkDeepMind math reasoning, IMO gold

Related

Referenced by

Sources