Microsoft MAI family of AI models (reasoning, coding, voice, images)
Key Questions
What models are included in Microsoft's new MAI family?
The MAI family includes MAI-Thinking-1 as its first reasoning model, MAI-Code-1-Flash (137B MoE coding model achieving 51% on SWE-bench Pro), plus image, transcription, and voice models. They were announced at Build 2026.
How does MAI-Thinking-1 differ from prior Microsoft AI efforts?
MAI-Thinking-1 is Microsoft's first advanced reasoning model trained cleanly without distillation. It matches leading benchmarks and signals greater independence from OpenAI.
What does the MAI family indicate about Microsoft's AI strategy?
The releases mark direct competition with GPT-5.5, Gemini, and Claude while reducing reliance on OpenAI partnerships. MAI-Code-1-Flash targets coding performance specifically.
Where were the MAI models announced?
Microsoft announced the MAI family at Build 2026 alongside updates on platforms and developer tools. The focus includes reasoning, voice, coding, and image capabilities.
How does MAI-Code-1-Flash perform on coding benchmarks?
MAI-Code-1-Flash scores 51% on SWE-bench Pro as a 137B MoE model. It contributes to Microsoft's push for in-house coding and reasoning excellence.
Microsoft launches MAI-Thinking-1 (first reasoning model, clean training, no distillation, matches leading benchmarks), MAI-Code-1-Flash (137B MoE coding model, 51% SWE-bench Pro), plus image, transcription, voice models. Signals Microsoft's independence from OpenAI and direct competition with GPT-5.5, Gemini, Claude. Announced at Build 2026.