Microsoft MAI-Thinking-1 Reasoning Model Achieves 97% on AIME Math Benchmark
Tags AI · Infrastructure · Enterprise
Microsoft unveiled its first independently-developed reasoning model, MAI-Thinking-1, achieving 97% accuracy on the AIME 2025 mathematics benchmark. The model uses a mixture-of-experts architecture with 1 trillion total parameters and 35 billion active per token, trained on 30 trillion pretraining tokens with over 50% code data. Available via Microsoft Foundry private preview and partner platforms including Fireworks AI and OpenRouter. This represents Microsoft's first foundation model built entirely in-house rather than through partnerships with OpenAI or DeepSeek.
Technical significance
Microsoft's first independently-built reasoning model demonstrates competitive performance against Anthropic's Claude Sonnet 4.6 and DeepSeek V3.2. The mixture-of-experts approach with 35B active parameters reduces inference costs while maintaining performance, signaling a shift toward specialized reasoning models for enterprise applications. Available through multiple partner platforms, reducing vendor lock-in concerns.