Microsoft unveils MAI-Voice-1 and MAI-1-preview, two in-house artificial intelligence models

Microsoft AI has introduced its first two models developed entirely in-house, marking a key step in building its own technological foundation in artificial intelligence. The first, MAI-Voice-1, is a voice generation system capable of producing expressive and natural audio. This model is integrated into Copilot Daily and Podcasts, and can also be tested in Copilot Labs, where demonstrations include storytelling and creative applications such as guided meditations or interactive stories. It can generate one minute of audio in less than a second using a single GPU, reflecting high technical efficiency compared to other systems currently available.

The second announcement is MAI-1-preview, the first foundational model trained entirely by Microsoft AI. It is a mixture-of-experts model trained and fine-tuned on approximately 15,000 NVIDIA H100 GPUs, designed to follow instructions and respond to everyday queries. The model is currently available on the LMArena community platform for public evaluation and will gradually be deployed in selected Copilot text functions in the coming weeks. Early access is also offered via an API to a limited group of trusted testers.

Microsoft has stated that this development is part of a broader approach that combines in-house models with partner solutions and open-source contributions, aiming to deliver results tailored to millions of daily interactions. The company anticipates that the future of its AI platform will rely on an orchestration of specialized models, each serving different uses and needs.

With the launch of its new GB200 cluster, the company now has the infrastructure required to continue developing more advanced models and expanding their applications in Copilot and other products. These announcements mark a significant step in Microsoft AI’s strategy to consolidate its position in the development of foundational and applied systems at a global scale.

Key points

• Microsoft AI presents its first two in-house models developed entirely internally.
• MAI-Voice-1 can generate one minute of expressive, natural audio in less than a second using a single GPU.
• The voice model is integrated into Copilot Daily, Podcasts, and available in Copilot Labs for creative demonstrations.
• MAI-1-preview is the first foundational mixture-of-experts model fully trained by Microsoft AI on 15,000 NVIDIA H100 GPUs.
• MAI-1-preview is publicly tested on LMArena and will be gradually deployed in Copilot in the coming weeks.
• Microsoft offers early access to MAI-1-preview via an API for trusted developers.
• The company maintains a hybrid approach combining in-house models with partner solutions and open-source software.
• Microsoft AI plans a future orchestration of specialized models for different uses and specific needs.
• The new GB200 cluster provides the infrastructure needed to develop more advanced models.
• These launches consolidate Microsoft AI’s position in the development of foundational models on a global scale.