# MAI-Thinking-1

> **TL;DR.** MAI-Thinking-1 is Microsoft's first in-house flagship reasoning model — a 35B active-parameter sparse Mixture of Experts system with a 256k-token context window, trained entirely on 30T tokens of licensed human data without distilling any third-party model.

- **Category:** AI / Language Models / Enterprise
- **Stage:** nascent
- **Age:** 1 days
- **Origin date:** 2026-06-02
- **First detected:** 2026-06-03
- **Canonical URL:** https://earlyterms.com/term/mai-thinking-1
- **Sources:** 7 primary URLs

## Definition

MAI-Thinking-1 is Microsoft's first in-house flagship reasoning model — a 35B active-parameter sparse Mixture of Experts system with a 256k-token context window, trained entirely on 30T tokens of licensed human data without distilling any third-party model.

Announced at [Microsoft Build 2026 on June 2](https://microsoft.ai/news/introducing-mai-thinking-1/) alongside six other MAI models, it scored 97.0% on AIME 2025 and was preferred over Anthropic Claude Sonnet 4.6 in 1,276 blind human evaluations. Available in private preview via Azure AI Foundry and GitHub Models, it signals Microsoft's pivot away from OpenAI dependency toward self-sufficient AI infrastructure.

## Analogy

Think of it as the moment a major automaker stopped buying engines from a single supplier and built its own.

## Why it's emerging now

Microsoft amended its OpenAI partnership in April 2026, ending exclusive licensing obligations. Build 2026 on June 2 delivered the first tangible proof: seven in-house MAI models led by MAI-Thinking-1, trained on 30T tokens of licensed human-generated data with zero third-party distillation — a direct counter-positioning to the DeepSeek data-provenance controversy that spooked enterprise procurement in early 2025.

## Related terms

- *related:* MAI-Code-1-Flash
- *related:* Azure AI Foundry
- *competitor:* gpt-5-4
- *competitor:* claude-opus-4-8
- *competitor:* kimi-k2-6
- *competitor:* glm-5-1
- *competitor:* deepseek-v4
- *related:* agent-harness
- *parent:* agentic-frameworks
- *related:* managed-agents
- *related:* phi
- *related:* Mixture of Experts

## Sources

1. [Microsoft AI — MAI-Thinking-1 official introduction](https://microsoft.ai/news/introducing-mai-thinking-1/)
2. [Microsoft AI — Building a hill-climbing machine: seven new MAI models](https://microsoft.ai/news/building-a-hillclimbing-machine-launching-seven-new-mai-models/)
3. [MAI-Thinking-1: Building a Hill-Climbing Machine (technical paper)](https://microsoft.ai/wp-content/uploads/2026/06/main_20260602_2.pdf)
4. [TechTimes — MAI-Thinking-1: first in-house reasoning model, trained without OpenAI data](https://www.techtimes.com/articles/317631/20260602/microsoft-build-2026-mai-thinking-1-first-house-reasoning-model-trained-without-openai-data.htm)
5. [Neowin — Microsoft unveils MAI-Thinking-1 and MAI-Code-1 models](https://www.neowin.net/news/microsoft-unveils-mai-thinking-1-reasoning-and-mai-code-1-coding-models/)
6. [Microsoft Community Hub — New MAI models in Azure AI Foundry](https://techcommunity.microsoft.com/blog/azure-ai-foundry-blog/new-mai-models-in-microsoft-foundry-across-text-image-voice-and-speech/4524632)
7. [Hacker News — MAI-Thinking-1 (188 points, 78 comments)](https://news.ycombinator.com/item?id=48374362)

---
_Generated by EarlyTerms · https://earlyterms.com/term/mai-thinking-1_