Microsoft Built 7 AI Models From Scratch. Here's What That Actually Means.

At Build 2026, Microsoft announced 7 AI models built entirely in-house. Here's what they actually built.

At Build 2026 on June 2, the Microsoft AI team led by Mustafa Suleyman announced the MAI model family: 7 models covering reasoning, coding, images, voice, and transcription. Every one built entirely in-house, trained from scratch.

Built From Scratch — What That Actually Means

Most AI companies use distillation — training a smaller model to imitate a larger one. The problem is a copied model can never exceed the original. It has a ceiling built in. Microsoft skipped distillation entirely and used only clean, commercially licensed data throughout. No murky training sources, no legal gray areas. Suleyman called it a "hill-climbing machine": models improving through their own internal cycles, not riding someone else's work.

The 7 Models at a Glance

ModelWhat it doesStatus
MAI-Thinking-1Reasoning (35B active, 1T total params, 256K context)Private preview
MAI-Code-1-FlashCoding, 5B params, 51% SWE-Bench ProLive in Copilot + VS Code
MAI-Image-2.5Text-to-image + image editing, #3 on Arena.aiLive in PowerPoint, OneDrive
MAI-Image-2.5 FlashFaster, lower cost image variantLive
MAI-Transcribe-1.5Speech-to-text, 43 languages, 276x real-timeLive in Copilot, Teams
MAI-Voice-2Text-to-speech + voice cloning, 15 languagesLive
MAI-Voice-2 FlashLow-latency variant for real-time agentsComing soon

MAI-Thinking-1 scored 97% on AIME 2025 and 53% on SWE-Bench Pro. Human evaluators preferred it over Claude Sonnet 4.6 in blind tests. MAI models in Azure are priced 20 to 60% lower than comparable OpenAI models.

We'll Test These

MAI-Thinking-1 is in private preview and MAI-Code-1-Flash is already live via Microsoft Foundry, OpenRouter, Fireworks AI, and Baseten. As access opens up, we'll run our own benchmarks — the same enterprise test suites we used for the Gemma family — and post the real numbers here.

Comments

Leave a comment