Microsoft used its annual Build developer conference to announce seven new AI models developed entirely in-house under the MAI brand. The launch, announced by Microsoft AI chief Mustafa Suleyman, spans reasoning, coding, image generation, transcription, and voice. It is the broadest model release in the company's history and marks a clear signal that Microsoft is building its own AI capabilities independent of its relationship with OpenAI.
The headline model is MAI-Thinking-1, Microsoft's first purpose-built reasoning model. All seven models were trained from scratch on clean, commercially licensed data, with no distillation from any third-party AI lab. Suleyman framed the launch as the debut of what he calls a 'hill-climbing machine', a shared training infrastructure designed to keep Microsoft's models competitive as global compute resources scale dramatically over the coming years.
MAI-Thinking-1
MAI-Thinking-1 is Microsoft AI's flagship reasoning model and the centrepiece of the Build announcement. It is a sparse Mixture-of-Experts model with 35 billion active parameters and approximately one trillion total parameters, an architecture that delivers a smaller inference footprint than dense models of comparable capability. Microsoft says it matches Claude Opus 4.6 on the SWE-Bench Pro software engineering benchmark and reaches 97.0 percent on the AIME 2025 mathematics benchmark and 94.5 percent on AIME 2026, placing it among the strongest models in its weight class on both coding and maths tasks.
In human preference testing, Microsoft conducted 1,350 blind side-by-side evaluations using professional raters from Surge, covering single-turn and multi-turn conversations across a wide range of tasks. Users preferred MAI-Thinking-1's responses over those from Claude Sonnet 4.6, a finding Microsoft says demonstrates that benchmark performance is translating into genuine real-world usefulness. The model supports a 256,000-token context window (enough to fit a 600-page document), function calling, and developer instructions, and is compatible with the Chat Completions API. It is currently available in private preview on Microsoft Foundry, with a public preview on the MAI Playground coming soon.
The other six models
MAI-Code-1-Flash is a lightweight agentic coding model with five billion parameters, positioned as comparable to Haiku but at lower cost. It is deeply integrated into GitHub Copilot and VS Code and is designed for engineering teams that need fast, continuous coding assistance without the overhead of a larger model.
MAI-Image-2.5 handles both text-to-image generation and image editing from a single model. Microsoft says it debuted at No. 2 on the Arena image editing leaderboard, surpassing the score of Nano Banana Pro. An ultra-efficient Flash variant, MAI-Image-2.5 Flash, is also available, offering the same capabilities at lower inference cost.
MAI-Transcribe-1.5 is Microsoft's transcription model, which the company describes as the most accurate in the world. It operates five times faster than competing models and supports domain-specific terminology across 43 languages. It leads on the FLEURS and Artificial Analysis accuracy benchmarks.
MAI-Voice-2 provides high-quality, natural-sounding speech synthesis across 15 languages and can adapt to a new voice from a short audio sample. The model ships with safeguards designed to prevent misuse. A lower-cost, ultra-efficient variant, MAI-Voice-2-Flash, has been announced and is coming soon.
The hill-climbing machine
All seven models are built on a shared foundation Microsoft describes as its Hill-Climbing Machine: a co-designed training pipeline built to improve continuously as the company applies more compute, better data, and sharper evaluation methods. Microsoft said it co-designs its models with its in-house Maia 200 silicon accelerators and has already seen a 1.4x efficiency improvement from this integration. Its next-generation GB200 compute cluster is now operational. Suleyman projected that the compute available to train frontier models will increase by a further thousand-fold over the next three years.
Frontier Tuning
Alongside the models, Microsoft introduced Frontier Tuning, a system that lets enterprises build custom versions of MAI models by training them on their own workflows and proprietary data using reinforcement learning in dedicated environments. The resulting model stays within the enterprise's own infrastructure and embodies the organisation's institutional knowledge. Microsoft said a MAI model tuned for Excel tasks matches GPT 5.4 in performance while running up to ten times more efficiently. When tuned to McKinsey's enterprise standards, MAI achieved the highest win rate of any model tested at approximately ten times lower cost.
A healthcare model with Mayo Clinic
Microsoft also announced a collaboration with Mayo Clinic to co-develop a frontier AI model for clinical use. The model will be trained on Mayo Clinic's de-identified patient data and longitudinal clinical insights, combined with Microsoft's foundational AI capabilities. It will first be deployed within Mayo Clinic's own systems for applications including earlier diagnosis and treatment planning, before being made available to other healthcare organisations via Azure Foundry. Ownership of the model will remain with Mayo Clinic.
Availability
All seven models are available or coming soon on Microsoft Foundry. They will also be accessible to developers via OpenRouter, Fireworks, and Baseten. Microsoft said that for the first time, developers will be able to tune the model weights directly. The company described the broader ambition behind the launch as building toward what it calls Humanist Superintelligence, AI systems that serve people and organisations, remain subordinate to human oversight, and do not replace the humans they work with.
