Microsoft used its Build keynote on Tuesday to roll out seven in-house AI models, with MAI-Thinking-1 as the headline. It's the company's first reasoning model, and the pitch is independence: trained from scratch, no distillation from anyone else's models, on what Microsoft calls clean, commercially licensed enterprise data.
The specs are mid-sized by design. A Mixture of Experts setup with 35 billion active parameters, a 256K context window, built for low token cost. The interesting comparison is who Microsoft picked to measure against. The model page says blind human raters preferred MAI-Thinking-1 over Claude Sonnet 4.6, and that it matches Claude Opus 4.6 on coding via SWE Bench Pro. Those ratings came from Surge, Microsoft's own rating partner, and the company hasn't published numbers for the side-by-side preference claim. So treat the win as self-reported for now.
The rest of the family fills in modalities. MAI-Image-2.5 for image generation, MAI-Transcribe-1.5 across 43 languages, MAI-Voice-2 with 15-plus added languages, and MAI-Code-1, an inference-tuned coding model.
MAI-Code-1 is the one developers can touch today. It's live in GitHub Copilot and VS Code right now. MAI-Thinking-1 stays in private preview on Foundry, open only to early access. Microsoft also said the MAI models will land on Fireworks AI, Baseten, and OpenRouter.
Bottom Line
MAI-Thinking-1 runs 35 billion active parameters with a 256K context window, but it's locked in private preview while MAI-Code-1 ships to Copilot today.
Quick Facts
- Model: MAI-Thinking-1, Microsoft's first reasoning model
- 35 billion active parameters, Mixture of Experts architecture
- 256K-token context window
- Preferred over Claude Sonnet 4.6 in blind tests (company-reported, via Surge)
- MAI-Code-1 live in GitHub Copilot and VS Code; MAI-Thinking-1 in private preview
