Microsoft AI unveils its first in-house models, MAI-Voice-1 and MAI-1-preview

Background

On 28 August 2025 Microsoft AI announced its first two in-house models, MAI-Voice-1 and MAI-1-preview.^[1] Coverage framed the pair as Microsoft's first step toward reducing its dependence on OpenAI.^[3] The two models are different kinds of system. One generates speech. The other is a general language model that generates text.

What happened

MAI-Voice-1 is a speech generation model, a system that turns written text into spoken audio. Microsoft said it can produce a minute of audio in under a second on a single GPU, the graphics processor that most AI systems run on.^[1] The model had already been put to work inside Microsoft's own products, where it powered Copilot Daily and Podcasts.^[1]

MAI-1-preview is a foundation model, meaning a large general-purpose model that is first trained on a broad body of data and then adjusted for particular uses. It is built as a mixture-of-experts (MoE) model. In this design the network is split into many specialised sub-networks, and for each input only a few of them are switched on, so the model can be large without using all of its parameters every time. Microsoft said the model was pre-trained and post-trained on about 15,000 NVIDIA H100 GPUs.^[1]

Key details

MAI-1-preview went into public testing on LMArena, a platform where people compare the answers of different models in side-by-side votes. On the text leaderboard there it debuted in 13th place.^[1][2] Microsoft said the model was due to reach some Copilot text features in the following weeks.^[1][2]

Microsoft also said that its next-generation GB200 cluster, a group of NVIDIA GB200 systems linked together to train and run large models, was already operational.^[1] For a company that trains its own models, the availability of this kind of hardware sets how large and how fast those models can be built.

Impact

The release showed that Microsoft had started building its own models for its products, alongside the models it offered from other companies. The significance lies less in any single leaderboard placement than in the direction it signalled. Coverage tied the models to Microsoft's wish to rely less on one outside supplier.^[3]

For readers following the field, the announcement is a useful marker. It shows a large platform company moving from being a customer of foundation models toward training them, with a speech model and a text model released at the same time and tested in public.^[1][3]