MAI-1-preview

What MAI-1-preview is

MAI-1-preview is a text foundation model built by Microsoft AI and announced on 28 August 2025. Microsoft described it as the group's first foundation model trained end to end in-house: pre-training and post-training were both carried out by Microsoft itself rather than adapted from a model built elsewhere.^[1][3] A foundation model in this sense is a general-purpose model trained on large amounts of text that can later be tuned or prompted for many different tasks.

How it was built

The model uses a mixture-of-experts architecture. In such a model the network contains many separate expert sub-networks plus a routing mechanism that sends each piece of input to only a small subset of them, so the model can hold far more parameters in total than it activates for any single token. Microsoft said MAI-1-preview was pre-trained and post-trained on about 15,000 NVIDIA H100 GPUs.^[1] It did not publish the parameter count, the context length, or detailed benchmark results.^[2] That leaves the model's size and capability profile only partly described in public.

Release and early results

Rather than a full product launch, Microsoft opened MAI-1-preview for public testing on LMArena, a platform where visitors are shown two anonymous model responses to the same prompt and pick the better one, with the votes feeding a public leaderboard. There the model initially ranked 13th on the text leaderboard.^[1][2] Such a placement reflects users' pairwise preferences over prompts they choose themselves rather than a fixed benchmark suite, so it says how the outputs were received rather than giving a controlled measurement of capability.

Microsoft said the model would roll out to certain text use cases in Copilot over the following weeks, and that it offered API access to trusted testers.^[1] Microsoft AI described MAI-1-preview as offering a glimpse of future offerings inside Copilot.^[1]

Announced alongside MAI-Voice-1

MAI-1-preview was announced together with MAI-Voice-1, a model Microsoft described as delivering high-fidelity expressive audio for single- and multi-speaker scenarios.^[1] In the same announcement Microsoft said its next-generation GB200 cluster was operational, adding computing capacity for training work.^[1] Announcing a text model, a speech model and new hardware together framed the release as a step in building out a model family rather than a single product.

How Microsoft framed it

Microsoft said it would orchestrate a range of specialised models serving different user intents rather than relying on one model for everything.^[1] In that framing MAI-1-preview is one component of a larger system, with different models handling different kinds of requests, instead of the single engine behind Copilot.

Where it sits

The release marked Microsoft's move to build its own frontier models alongside its continuing use of OpenAI models.^[3] Its successor as Microsoft AI's flagship text model was MAI-Thinking-1, released in June 2026.^[3]