What Is MAI-Thinking-1? Microsoft's First In-House Reasoning Model
For four years, nearly every AI-powered click inside Copilot was, underneath, a call to a model Microsoft didn't build. At Build 2026, that started to change. Microsoft AI unveiled seven in-house MAI models, and MAI-Thinking-1 is the headline: a reasoning model trained from scratch, on Microsoft's own data and infrastructure, with no distillation from OpenAI or anyone else. Here's what it actually is, and what "Microsoft finally builds its own models" really means for the products you use.
MAI-Thinking-1 was unveiled at Microsoft Build 2026 alongside six other in-house MAI models.
Built From Scratch, No Distillation
MAI-Thinking-1 is Microsoft AI's first in-house reasoning model, unveiled June 2, 2026 at Microsoft's Build developer conference. It's a mixture-of-experts model with about 35 billion active parameters and roughly 1 trillion total parameters, currently in private preview on Microsoft Foundry. The model's defining claim is how it was trained: Microsoft says MAI-Thinking-1 was built "without distillation from third-party models" — meaning it wasn't fine-tuned on outputs from GPT, Claude, or any other lab's system, but trained end-to-end on Microsoft's own clean, traceable, enterprise-grade data.
It shipped alongside six other MAI-branded models spanning coding, image generation, voice synthesis, and transcription — most notably MAI-Code-1-Flash, a smaller coding model already rolling out as a selectable option in the GitHub Copilot model picker for VS Code.
Microsoft's Three-Pillar Philosophy
Learned, not inherited
Microsoft argues that distilled models are tied to their teacher's design choices and struggle to adapt to new situations — MAI-Thinking-1 was trained to learn tasks directly instead.
Clean, traceable data
Trained from the ground up on data Microsoft can account for, which the company frames as essential for understanding and improving model behavior over time.
Full-stack self-sufficiency
Co-designed with Microsoft's own accelerators and trained on in-house reinforcement learning infrastructure, rather than renting another lab's training stack.
The "Hill-Climbing Machine"
Microsoft frames MAI-Thinking-1 as one output of a repeatable training pipeline built to absorb better data, stronger rewards, and more compute continuously over time.
Independent human raters on the Surge evaluation platform reportedly preferred MAI-Thinking-1 over Claude Sonnet 4.6 in blind side-by-side comparisons for overall quality across single- and multi-turn tasks — a genuinely third-party evaluation method, though the specific comparison model (Sonnet 4.6, not a more recent release) and the full methodology are worth checking directly in Microsoft's technical report before drawing conclusions.
Less Dependence, Not a Clean Break
MAI-Code-1-Flash is already available as a selectable model inside the GitHub Copilot model picker.
| MAI-Thinking-1 | MAI-Code-1-Flash | |
|---|---|---|
| Role | Flagship reasoning model | Small-tier coding model |
| Status | Private preview, Microsoft Foundry | Live in GitHub Copilot model picker |
| Training data | No third-party distillation | Trained on Copilot's own agentic harness |
| Availability | Also listed on OpenRouter, Fireworks, Baseten | Rolling into VS Code auto-picker |
This isn't Microsoft dropping OpenAI. The company's roughly $13 billion partnership with OpenAI and its infrastructure relationship with Anthropic both continue — Copilot's model picker is adding MAI options alongside existing ones, not replacing them outright. The strategic signal is independence: after years of running almost entirely on licensed models, Microsoft now has credible frontier-adjacent reasoning and coding models of its own, giving it leverage it didn't previously have in pricing and product decisions.
Why Building In-House Models Actually Saves Microsoft Money
The financial motivation behind MAI-Thinking-1 is worth spelling out beyond the "independence" framing, because it explains why a company already paying billions for access to OpenAI's models would also invest heavily in building its own. Every Copilot query routed through a licensed third-party model carries a per-token cost Microsoft pays on top of its own infrastructure spend — a cost that scales directly with Copilot's user base and usage volume. A first-party model running on Microsoft's own Azure infrastructure and custom accelerators removes that per-query licensing cost entirely for the queries it handles, even though the upfront investment in training and infrastructure is substantial. At Copilot's scale, shifting even a meaningful minority of queries from a licensed model to an in-house one can represent significant savings over time, which is a large part of why every major AI lab with a consumer product at scale has pursued the same in-house model strategy.
There's a second, less obvious benefit: pricing leverage in negotiations with external providers. A company with no credible in-house alternative has little room to negotiate licensing terms, since switching away from a provider isn't a realistic option. Once Microsoft has models that are "good enough" for a meaningful share of Copilot's traffic, it gains real negotiating leverage with OpenAI and Anthropic for the traffic it continues to route to them — a dynamic that likely factors into the strategy as much as the technical independence Microsoft emphasizes publicly.
Who MAI-Thinking-1 Is For
- Genuinely trained without third-party distillation — a real architectural claim, not marketing
- Strong reported performance for a mid-sized, 35B-active-parameter model
- Independent human-preference evaluation via Surge, not just self-reported benchmarks
- Backed by Microsoft's enterprise distribution across Foundry, Copilot, and Azure
- Still in private preview — not yet broadly available to test
- Headline benchmark figures (AIME scores) are Microsoft-reported
- Compared against Claude Sonnet 4.6, not necessarily the current frontier model at review time
- Microsoft's first public reasoning model, with limited independent long-term track record
Six Other Models, One Strategy
MAI-Code-1-Flash
A small-tier coding model already live in the GitHub Copilot model picker, trained directly on Copilot's own file-editing tools and terminal integrations.
MAI-Image-2
Microsoft's in-house image generation model, part of the same April 2026 wave of Azure-side model launches that preceded Build's coding and reasoning releases.
MAI-Voice-1
A voice synthesis model shipped alongside MAI-Image-2, aimed at Microsoft's transcription and narration features across Office and Teams.
Transcription models
Additional MAI models handle meeting transcription, feeding into Teams and Microsoft 365 Copilot's note-taking and summarization features.
Taken together, Microsoft's own framing is that no single MAI model is meant to be a general-purpose chatbot competing head-on with ChatGPT. Each is fitted to a specific tool surface — reasoning for Foundry and complex enterprise tasks, coding for Copilot's agentic harness, voice and transcription for Office — reflecting a broader 2026 industry trend toward models built for a particular workflow rather than sold as one-size-fits-all systems.
Frequently Asked Questions
What is MAI-Thinking-1?
Microsoft AI's first in-house reasoning model, unveiled at Build 2026, trained from scratch without distillation from any third-party model.
Does this mean Microsoft is dropping OpenAI?
No — MAI models are being added as additional options in Copilot and GitHub Copilot alongside existing OpenAI and Anthropic models.
How does it perform compared to other models?
Microsoft reports 97.0% on AIME 2025 and says human raters preferred it over Claude Sonnet 4.6 in blind tests — figures worth verifying independently.
More AI Model Guides, Written Clearly
Browse our full library of AI model explainers, updated as new releases land.
More from the creator
Explore two other projects from the same creator.