Redmond, Washington, United States— Microsoft revealed its first AI models developed entirely in-house: MAI-Voice-1 AI and MAI-1-preview.
The MAI-Voice-1 model can generate a one-minute audio clip in less than a second using a single graphics processing unit. Microsoft is already using it for features such as Copilot Daily, which allows an AI virtual presenter to read top news stories and produce podcast-style discussions, according to The Verge.
Users can try the model through Copilot Labs, where they can enter text, and adjust the AI’s voice and speaking style.
The MAI-1-preview model was trained on roughly 15,000 Nvidia H100 GPUs and is designed for users who need AI capable of following instructions and providing helpful responses to everyday queries.
Microsoft has begun publicly testing the model on the benchmarking platform LMArena.
Mustafa Suleiman, head of Microsoft’s AI division, said the company’s internal models are not intended for enterprise use. “Our focus is on building models that genuinely function as consumer companions,” he said.
Microsoft plans to deploy MAI-1-preview in its Copilot smart assistant, which currently relies on large language models from OpenAI. The company said it aims to coordinate specialized models for different user goals and use cases to unlock significant value.
