The Declaration: Humanist Superintelligence
Microsoft AI CEO Mustafa Suleyman opened the Build 2026 keynote with a simple statement: Microsoft's goal is Humanist Superintelligence — AI designed to serve people, not replace them. The MAI model family is the first concrete step toward that mission.
More significant than the models themselves is what Suleyman emphasized about how they were built: "We don't distill from other labs and we don't rely on opaque data. Our datasets are clean, traceable, and enterprise-grade." This is a direct signal that Microsoft — which has been a major Azure reseller of OpenAI models — is now building its own complete AI stack.
The Seven Models, Explained
MAI-Image-2.5 & Flash
A multimodal image model supporting both text-to-image generation and precise image editing. It ranks #2 on the Arena leaderboard for image editing, surpassing Google's Nano Banana Pro. Already live in PowerPoint and rolling out to OneDrive, with availability on Microsoft Foundry for developers. The Flash variant optimizes for production-scale cost and latency.
MAI-Transcribe-1.5
The headline claim: the world's best transcription model, with state-of-the-art word error rate (WER) across 43 languages on the FLEURS multilingual benchmark. It transcribes an hour of audio in under 15 seconds — up to 5× faster than Gemini 3.1 and GPT-4o-Transcribe at comparable accuracy. Keyword Biasing allows users to supply domain-specific terminology, reducing WER by up to 30%. Integrating into Copilot, Teams, GitHub, and Dynamics 365 Contact Centre.
MAI-Voice-2 & Flash
High-quality speech generation across 15 languages, with the ability to adapt to a speaker's voice from a short audio sample. Voice-2-Flash is optimized for ultra-low-latency real-time voice agents.
MAI-Thinking-1
Microsoft AI's first dedicated reasoning model. It reaches 97.0% on AIME 2025 and 94.5% on AIME 2026, demonstrating strong mathematical and scientific reasoning competitive with models in its weight class. In blind human side-by-side evaluations, it outperforms Claude Sonnet 4.6. Currently available in private preview on Microsoft Foundry.
MAI-Code-1-Flash
At just 5B active parameters, MAI-Code-1-Flash achieves 51.2% on SWE-Bench Pro — 16 percentage points ahead of Claude Haiku 4.5 (35.2%) on the same benchmark. It solves harder problems with up to 60% fewer tokens on SWE-Bench Verified, reducing latency and cost simultaneously. Adaptive Solution Length Control automatically calibrates response depth to task complexity. Rolling out today as a selectable model in VS Code GitHub Copilot.
MAI models are available via Microsoft Foundry, as well as OpenRouter, Fireworks AI, and Baseten. For the first time, developers can fine-tune model weights directly through Foundry. MAI-Code-1-Flash requires no setup for GitHub Copilot users in VS Code — it appears in the model picker automatically as the rollout progresses.
The Full MAI Family at a Glance
| Model | Category | Key Metric | Where Available |
|---|---|---|---|
| MAI-Image-2.5 | Image gen + editing | Arena image editing #2 | PowerPoint, OneDrive, Foundry |
| MAI-Image-2.5-Flash | Image (efficient) | Cost/speed optimized | Foundry |
| MAI-Transcribe-1.5 | Speech-to-text | Best WER in 43 languages | Copilot, Teams, GitHub, Foundry |
| MAI-Voice-2 | Text-to-speech | 15 languages, voice adapt | Copilot |
| MAI-Voice-2-Flash | Voice (low latency) | Ultra-low latency | Voice agents |
| MAI-Thinking-1 | Reasoning | AIME 2026: 94.5% | Foundry (private preview) |
| MAI-Code-1-Flash | Coding | SWE-Bench Pro: 51.2% | GitHub Copilot, VS Code |
Alongside the MAI model family, Suleiman announced that Microsoft and Mayo Clinic are co-creating a frontier AI model for healthcare. The collaboration combines Mayo Clinic's world-leading clinical expertise and de-identified longitudinal clinical data with Microsoft's foundational AI capabilities — marking MAI's first major domain-specific vertical application.
- Microsoft AI launched 7 new MAI models at Build 2026 spanning image, transcription, voice, reasoning, and coding.
- All models trained on proprietary Microsoft data only — no distillation from OpenAI or third-party models.
- MAI-Code-1-Flash: 5B params, SWE-Bench Pro 51.2%, live in GitHub Copilot and VS Code immediately.
- MAI-Transcribe-1.5: best-in-class WER across 43 languages, 5× faster than comparable models.
- Available on Foundry, OpenRouter, Fireworks, Baseten — weight fine-tuning permitted for the first time.