Microsoft is taking another big step in its push to build more capable in-house AI with MAI-Thinking-1, its first reasoning model from the Microsoft AI team.
The company says the model was built from scratch, rather than distilled from models developed by other AI labs. MAI-Thinking-1 is designed for complex reasoning, mathematics, software engineering, and enterprise workloads—and it is now available through Microsoft Foundry in private preview.
That makes this more than just another model launch. It is an important milestone for Microsoft’s growing MAI model family.
Our first reasoning model, MAI-Thinking-1, is built from scratch. Now available in Microsoft Foundry. Kudos to the team! More below. pic.twitter.com/7NWith33MS
— Mustafa Suleyman (@mustafasuleyman) August 12, 2026
MAI-Thinking-1 Was Built From the Ground Up
One of the most interesting things about MAI-Thinking-1 is Microsoft’s approach to training it.
Microsoft says the model was trained without distillation from third-party models, using clean, traceable and enterprise-grade data. The company wants its models to learn capabilities directly rather than simply inheriting intelligence from another model.
That approach is part of Microsoft’s broader “Hill-Climbing Machine” strategy—a training and development pipeline designed to continuously improve models through better data, rewards, environments and compute.
For Microsoft, building this capability in-house is becoming increasingly important as it looks to compete with the likes of OpenAI, Anthropic, Google and other frontier AI developers.
A Surprisingly Compact Reasoning Model
Don’t let the model’s size fool you.
MAI-Thinking-1 is a Mixture-of-Experts (MoE) model with around 1 trillion total parameters and 35 billion active parameters. Because only parts of the model are activated for a given task, Microsoft can deliver substantial capabilities without requiring the inference footprint of a much larger dense model.
Microsoft is clearly targeting the sweet spot between capability and cost.
The model supports a 256K-token context window, making it suitable for long documents, complex instructions, extended agent traces and other enterprise workloads where context matters.
Strong Coding and Math Performance
Reasoning is where MAI-Thinking-1 is expected to make its biggest impression.
Microsoft says the model is competitive with leading models on software engineering benchmarks and matches Claude Opus 4.6 on SWE-Bench Pro, despite having a much smaller active parameter footprint.
Its mathematics performance is also notable. Microsoft reports scores of 97.0% on AIME 2025 and 94.5% on AIME 2026.
For developers, that combination could be particularly interesting. A model that can reason through difficult problems while also understanding, modifying, testing and debugging code is a natural fit for increasingly autonomous AI coding agents.
Humans Preferred It in Some Blind Tests
Benchmarks aren’t the only signal Microsoft is highlighting.
In a blind evaluation involving 1,276 tasks, professional raters working with Surge preferred MAI-Thinking-1 over Claude Sonnet 4.6 across the tested single-turn and multi-turn tasks.
Of course, these are Microsoft’s reported evaluation results, so they shouldn’t be treated as a universal ranking of AI models. Still, human preference testing is useful because benchmark scores don’t always translate directly into a better everyday experience.
Built for Enterprise AI
Microsoft isn’t positioning MAI-Thinking-1 as a research experiment alone.
The model is designed for enterprise deployment, with support for long-context workloads, function calling and developer instructions. Microsoft Foundry also provides enterprise security, compliance and managed infrastructure around its AI models.
This could make MAI-Thinking-1 particularly interesting for organizations building AI agents that need to work with internal documents, codebases, business processes and other proprietary information.
Microsoft also says its MAI models are designed to support deployment with first-party managed compute, granular quota and zero data egress in Foundry.
MAI-Thinking-1 Is Just the Beginning
Perhaps the bigger story is what this means for Microsoft AI.
MAI-Thinking-1 arrived alongside a broader family of Microsoft-built models covering reasoning, coding, image generation, voice and transcription. Microsoft is clearly trying to establish its own model stack rather than relying entirely on external AI providers.
And that strategy could become increasingly important as AI moves from simple chatbots toward coding agents, research systems and autonomous workflows.
For now, MAI-Thinking-1 gives Microsoft a serious first-party reasoning model to build upon.
The Bottom Line
MAI-Thinking-1 is a significant milestone for Microsoft AI.
It is Microsoft’s first reasoning-focused model, built from scratch, with a relatively efficient MoE architecture, a 256K context window, strong reported math and coding results, and a clear focus on enterprise deployment.
The model is currently available through Microsoft Foundry in private preview, with broader availability planned.
Microsoft’s reasoning ambitions are clearly just getting started—and MAI-Thinking-1 could be the foundation for a much larger in-house AI push.
Stay tuned to WinCentral for more Microsoft AI, Windows, Copilot and emerging AI model news.
