Back to Blog
OpenAI

Microsoft MAI-Thinking-1: A Fully Homegrown Reasoning Model Scoring 97% on AIME 2025

At Build 2026 on June 2, 2026, Microsoft officially unveiled MAI-Thinking-1 — the company's first fully in-house reasoning model, built without relying on distillation from OpenAI or DeepSeek. Scoring 97.0% on AIME 2025, it marks Microsoft's formal entry as an independent player in foundation model development.

Model Specifications

SpecDetails
DeveloperMicrosoft
ArchitectureSparse Mixture of Experts (MoE)
Total Parameters~1 trillion (1T)
Active Parameters35 billion (35B)
Context Window256K tokens
Training Data30T tokens (50%+ code)
Training Hardware8,000 GB200 GPUs
Distillation❌ Fully in-house, no third-party distillation
Release DateJune 2, 2026 (Build 2026)

Benchmark Performance

BenchmarkMAI-Thinking-1Claude Sonnet 4.6DeepSeek V3.2Claude Opus 4.6
AIME 202597.0%95.6%93.1%99.8%
AIME 202694.5%
SWE-Bench Pro52.8%
LiveCodeBench v687.7%

On AIME 2025, MAI-Thinking-1 scored 97.0%, surpassing Claude Sonnet 4.6 (95.6%) and DeepSeek V3.2 (93.1%) — second only to Claude Opus 4.6 (99.8%).

It's worth noting that these are Microsoft's self-reported numbers and have yet to be independently reproduced and verified.

The Full MAI Model Family

Microsoft released seven MAI models at Build 2026:

ModelTypeDescription
MAI-Thinking-1ReasoningFlagship reasoning model (the focus of this article)
MAI-Code-1-FlashCodingFast coding model powering Copilot
MAI-Image-2.5Image GenerationImage generation model
MAI-Image-2.5 FlashImage GenerationFast image generation
MAI-Voice-2VoiceVoice model
MAI-Voice-2 FlashVoiceFast voice model
MAI-Transcribe-1.5TranscriptionSpeech-to-text model

Why This Matters

1. Microsoft Breaks Free from OpenAI Dependency

In April 2026, Microsoft and OpenAI revised their partnership agreement, allowing Microsoft to develop its own foundation models. The launch of MAI-Thinking-1 signals that Microsoft has officially transitioned from being an OpenAI technology distributor to an independent competitor.

2. Trained on a Massive GB200 GPU Cluster

MAI-Thinking-1 was trained on 8,000 GB200 GPUs — one of the largest known GB200 cluster training runs to date. GB200 is NVIDIA's latest flagship GPU, underscoring the enormous hardware investment Microsoft has made.

3. Over 50% Code in Training Data

With more than 50% of its training data consisting of code, it's no surprise that MAI-Thinking-1 delivers strong results on SWE-Bench Pro (52.8%) and LiveCodeBench (87.7%). For coding-intensive use cases, this model could be a highly competitive option.

Recommended Use Cases

Use CaseRecommended ModelRationale
Mathematical reasoning (AIME-style)MAI-Thinking-197.0% on AIME 2025, second only to Opus 4.6
Everyday coding assistanceMAI-Code-1-FlashCopilot-integrated, optimized for speed
Complex reasoning and architecture designClaude Opus 4.8Strongest all-around capability
Large-scale batch processingMAI-Thinking-1MoE architecture enables efficient inference
Data-sensitive environmentsMAI-Thinking-1Azure-hosted with enterprise-grade security

The Big Picture

The launch of MAI-Thinking-1 represents a fundamental shift in Microsoft's AI strategy.

Key takeaways:

  • First fully in-house reasoning model — no distillation from OpenAI or DeepSeek
  • 97.0% on AIME 2025 — beating Claude Sonnet 4.6, trailing only Opus 4.6
  • Trained on 8,000 GB200 GPUs — one of the largest GB200 clusters in existence
  • Seven MAI models launched simultaneously — spanning reasoning, coding, image generation, voice, and transcription
  • Microsoft is now an independent AI competitor — no longer just OpenAI's distribution partner

For enterprises, MAI-Thinking-1 introduces a new option: a Microsoft-backed, Azure-hosted reasoning model with enterprise-grade security. As Microsoft integrates MAI models into Copilot and Azure AI, their real-world impact will become increasingly clear.

Comments (0)

Share:XHatena

Post a Comment

Loading...