MAI-1-preview
Microsoft AI's first in-house foundation model, a mixture-of-experts system for everyday queries
MAI-1-preview is a mixture-of-experts language model built in-house by Microsoft AI and previewed in 2025. It is the division's first foundation model trained end-to-end, pre-trained and post-trained on roughly 15,000 NVIDIA H100 GPUs, aimed at instruction-following and helpful answers to everyday questions.
- Microsoft AI's first foundation model trained end to end
- Mixture-of-experts design, pre- and post-trained on ~15,000 NVIDIA H100 GPUs
- Public testing began on LMArena in 2025
- API access offered to trusted testers by application
- Announced alongside the MAI Voice 1 speech model
What it is
MAI-1-preview is a large language model developed inside Microsoft AI, the division that builds the company's own AI models. It is described as the first foundation model the team trained end to end, and it uses a mixture-of-experts design, where different sub-networks handle different inputs. Microsoft AI says it is aimed at consumers, with strengths in following instructions and answering ordinary questions helpfully. The company began public testing of the model on LMArena in 2025 and described it as an early look at what will later appear inside Copilot.
How it came about
Microsoft AI says it spent the previous year assembling a team and infrastructure to build purpose-built models. MAI-1-preview was announced alongside MAI Voice 1, a speech generation model. Public testing began on LMArena, a community platform for evaluating models, and the company also opened an application process for API access for trusted testers. Microsoft AI said it planned to roll the model out for some text uses within Copilot over the following weeks and to share more in the coming months.
Inside the model
The model is a mixture-of-experts architecture built in-house and trained end to end, meaning one team handled the whole pipeline rather than adapting an outside model. Both pre-training and post-training ran on about 15,000 NVIDIA H100 GPUs. Post-training covers the stages after initial training, such as teaching the model to follow instructions and give helpful replies. Microsoft AI has not published parameter counts, training data, context length or a detailed architecture description.
What it can do
Microsoft AI positions the model as strong at following instructions and responding helpfully to everyday consumer questions. It is intended as a foundation for future features inside Copilot. No benchmark scores, context window sizes or specific capability evaluations were published with the preview announcement. The company said it was collecting early feedback to learn where the model performs well and where it needs work.
How to use it
During the preview, the model could be tried on LMArena and was offered to trusted testers who applied for API access. Microsoft AI said it would roll it out for certain text use cases within Copilot over the following weeks. No downloadable weights or licence terms were announced. The company also said it would keep using models from its own team, partners and the open-source community in its products.
Why it matters
MAI-1-preview marks Microsoft AI's move to train its own foundation models rather than relying only on outside models. The company says it wants to orchestrate a range of specialised models for different user needs, and describes this release as a first step. It also said the model offers a glimpse of future offerings inside Copilot.
Questions people ask about MAI-1-preview
What is MAI-1-preview?
It is a large language model built in-house by Microsoft AI, the group that develops the company's own AI models. Microsoft AI describes it as its first foundation model trained end to end, using a mixture-of-experts design, and aimed at following instructions and answering everyday consumer questions.
How can I try it?
During the preview, the model was available for public testing on LMArena, a community platform for evaluating models. Microsoft AI also let trusted testers apply for API access. The company said it would roll the model out for certain text use cases within Copilot over the following weeks.
How large is it?
No parameter count has been published. Microsoft AI has said only that MAI-1-preview is a mixture-of-experts model and that both pre-training and post-training used roughly 15,000 NVIDIA H100 GPUs.
Is it open source?
No downloadable weights or licence terms were announced with the preview. Access was through LMArena, an application process for API access for trusted testers, and planned text features inside Copilot.
How does it relate to Copilot?
Microsoft AI described the model as an early look at future offerings inside Copilot and said it would be rolled out for certain text uses there. The company also said it would continue mixing its own models with partner and open-source models across its products.
Good reading and sources
There is no English Wikipedia article on MAI-1-preview yet, so the facts on this page come from the sources listed above: microsoft.ai/news/two-new-in-house-models/. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.
It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.