Microsoft AI · 2026-06-02 · major
Microsoft Launches Seven New MAI Models at Build 2026 — MAI-Thinking-1 Reasoning, MAI-Code-1-Flash 5B for Copilot, MAI-Image-2.5 With a Flash Variant, MAI-Voice-2 Across 15 Languages, MAI-Transcribe-1.5 Across 43
Mustafa Suleyman's MAI team rolled out its broadest model release yet: a first-party reasoner, a small coding model targeting Copilot, refreshed image and voice families, and a 5x-faster speech-to-text model trained without third-party distillation.
Microsoft AI's biggest first-party launch yet: a reasoner, a coding flash model, image + voice updates, and a speech-to-text engine claimed to be 5x faster than competing models.
Key specs
| Models launched | 7 |
|---|---|
| Mai code 1 flash params | 5B |
| Mai voice 2 languages | 15 |
| Mai transcribe 1.5 languages | 43 |
What is it?
Seven models out of Microsoft AI under the MAI brand: MAI-Thinking-1 (Microsoft's first dedicated reasoning model, medium-sized, text-only), MAI-Code-1-Flash (5B parameters, agentic coding, integrated with GitHub Copilot and VS Code), MAI-Image-2.5 plus a Flash variant (text-to-image and image editing), MAI-Voice-2 (text-to-speech across 15 languages, can adapt from short voice samples) with a Flash variant coming soon, and MAI-Transcribe-1.5 (speech-to-text across 43 languages). All sit under the 'hill-climbing machine' framing of iterative, internal model improvement.
How does it work?
Training relied on clean, in-house data with no distillation from third-party models, per the announcement. MAI-Thinking-1 is positioned to match leading reasoners on key software-engineering benchmarks and claims human-preference parity with Claude Sonnet 4.6. A custom-tuned MAI variant for Excel is said to match GPT-5.4 at roughly 10x lower cost. Distribution is through Azure AI Foundry, Open Router, Fireworks, and Baseten, with developers able to tune the weights themselves. MAI-Image-2.5 reportedly surpasses Google's Nano Banana Pro on Arena score; MAI-Transcribe-1.5 cites SOTA on FLEURS and Artificial Analysis.
Why does it matter?
Microsoft pays OpenAI billions a year for GitHub Copilot inference; rolling its own reasoning + coding model is the first concrete attempt to wean Copilot off that dependency. After the Microsoft-OpenAI partnership rewrite capped exclusivity and revenue share, this is the operational answer. MAI-Voice-2 multilingual + MAI-Transcribe-1.5's 43 languages also signal a real fight in the speech stack against ElevenLabs and Whisper-class systems.
Who is it for?
Enterprise developers on Microsoft stack, Copilot users, Foundry tenants, speech-AI buyers