Key takeaways
- Microsoft introduced seven new AI models aimed at reducing reliance on OpenAI and Anthropic.
- The models cover various functions including reasoning, coding, image generation, transcription, and voice.
- MAI-Image-2.5 excels in text-to-image generation and editing, outperforming previous benchmarks.
- MAI-Voice-2 offers multilingual speech generation, with a cost-effective Flash variant expected soon.
- Microsoft aims for advanced AI systems that enhance human capabilities rather than replace them.
At Microsoft Build 2026, Microsoft announced seven new in-house AI models spanning reasoning, coding, image generation, transcription and voice as it looks to reduce dependency on OpenAI and Anthropic.
The new model family includes MAI-Thinking-1, a reasoning model; MAI-Code-1-Flash for coding tasks; MAI-Image-2.5 and its Flash variant for image generation and editing; MAI Transcribe-1.5 for speech-to-text; and MAI-Voice-2 for speech generation.
Microsoft said MAI-Image-2.5 supports text-to-image generation and image editing, surpassing the Arena score of Nano Banana Pro. The company added that MAI-Voice-2 provides speech generation across 15 languages, while MAI-Voice-2-Flash, due to launch later, will offer similar capabilities at a lower cost.
The company added that the models will be available through Azure Foundry and developer platforms, including Hugging Face and GitHub.
Mustafa Suleyman, CEO of Microsoft AI, said Microsoft expects a further thousand-fold increase in AI training compute over the next three years, leading to more advanced capabilities and broader adoption of AI technologies.
“Our ultimate goal is what we call Humanist Superintelligence,” Suleyman said. “That means advanced AI systems designed to serve people and organisations, not replace them.”
Microsoft said MAI-Thinking-1 was trained without distillation from third-party models and matches leading systems on software engineering benchmarks. The company also said MAI-Code-1-Flash, a 5-billion-parameter model, is designed for integration with GitHub Copilot, Visual Studio Code and Microsoft’s software ecosystem.
Alongside the model launches, Microsoft introduced Frontier Tuning, a reinforcement learning approach that allows organisations to adapt models using their own workflows and operational data.
The company said a tuned MAI model for Excel matched the performance of GPT 5.4 while being up to 10 times more efficient. It also said a customised MAI model achieved the highest win rate in tests conducted to meet consulting firm McKinsey’s enterprise standards at roughly one-tenth the cost of competing models.
Microsoft also announced a collaboration with the Mayo Clinic to develop a healthcare-focused AI model. The model will combine Mayo Clinic’s clinical expertise and de-identified clinical data with Microsoft's AI technology. According to the company, the system will initially be deployed within Mayo Clinic's environment before becoming available to other organisations through Azure Foundry after validation.
The company said the healthcare model will be owned by Mayo Clinic and is intended to support clinical reasoning, diagnosis and treatment planning while maintaining patient trust and responsible data stewardship.
Microsoft said all MAI models were developed using internally built infrastructure, licensed datasets and the company's Maia 200 AI chip. The company reported a 1.4-times efficiency improvement from its hardware and software co-design efforts.
