Microsoft AI has released its first inference model, MAI-Thinking-1, along with six native new models, and introduced a private reinforcement learning service called "Frontier Fine-tuning."

According to Dyno Beat monitoring, Mustafa Suleyman, Head of Microsoft's AI division (Microsoft AI or MAI), announced at the Build 2026 developers' conference the launch of a new proprietary MAI native model family. The series includes a total of 7 models covering areas such as reasoning, programming, image, transcription, and speech. All models were trained from scratch by Microsoft without using any third-party models for knowledge distillation, and all datasets have been legally authorized. Microsoft stated that it is committed to building "Humanist Superintelligence," ensuring cutting-edge AI serves as a human assistant and is subject to human oversight. Additionally, Microsoft's deployed GB200 compute cluster is now fully operational to drive the continuous iteration of this model ecosystem.

The flagship reasoning model of the MAI family, MAI-Thinking-1, boasts 350 billion active parameters, utilizes a Mixture of Experts (MoE) architecture, and offers a 128K context window. In mainstream software engineering and mathematical reasoning evaluations such as SWE-bench Pro, this model has achieved performance equivalent to Claude Opus 4.6 and outperformed Claude Sonnet 4.6 in blind human assessments. For programming scenarios, MAI has introduced the intelligent agent programming model, MAI-Code-1-Flash, with 50 billion parameters, deeply integrated into GitHub Copilot and VS Code, providing performance similar to Claude Haiku at a lower inference cost. In the multimodal aspect, MAI-Image-2.5 and its Flash variant support high-precision text-to-image generation and image editing, surpassing Nano Banana Pro in image quality scores; in the speech and transcription domain, MAI has launched the state-of-the-art 43-language transcription model, MAI-Transcribe-1.5, with 5x faster speed compared to competitors, and the speech generation model, MAI-Voice-2, supporting 15 languages, emotional control, and zero-shot cloning, along with its Flash variant. These models are not only deployed on Azure AI Foundry but will also be listed on OpenRouter, Fireworks, and Baseten, with initial support for developers for fine-tuning weights. Microsoft also revealed that through joint software and hardware optimization of the models with the proprietary Maia 200 chip, a 1.4x computational efficiency improvement has been achieved.

In addition to the base model releases, Microsoft has launched the "Frontier Tuning" service based on Reinforcement Learning Environment (RLE). This service allows enterprises to perform custom training of MAI models in a fully controlled isolated environment ("Training Gym") using internal operation traces, decision sequences, and proprietary data. Tests have shown that custom models trained through Frontier Tuning experience significantly improved efficiency, with the Excel-optimized MAI model aligning in performance with GPT-5.4 but exhibiting 10x higher efficiency. The MAI model tailored for McKinsey achieved the highest win rate while reducing costs by nearly 10x. Furthermore, Microsoft announced a strategic partnership with the world-renowned healthcare institution, Mayo Clinic, to jointly develop a clinical reasoning large model based on Mayo's clinical data and Microsoft's AI infrastructure. Owned by Mayo Clinic, this model will initially be deployed internally at Mayo for early diagnosis and treatment plan design and later made available to other healthcare institutions through Azure AI Foundry.