MAI-1-preview
Last reviewed
Jun 8, 2026
Sources
8 citations
Review status
Source-backed
Revision
v1 · 1,546 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Jun 8, 2026
Sources
8 citations
Review status
Source-backed
Revision
v1 · 1,546 words
Add missing citations, update stale details, or suggest a clearer explanation.
MAI-1-preview is a large language model developed by Microsoft AI, the consumer artificial intelligence division of Microsoft led by Mustafa Suleyman. Announced on August 28, 2025, it was described by Microsoft as the company's first foundation model trained end to end entirely in-house, rather than being fine-tuned from a partner's system such as those of OpenAI or adapted from open-weight releases. MAI-1-preview is a mixture of experts text model that Microsoft said was pre-trained and post-trained on roughly 15,000 Nvidia H100 graphics processing units, and it debuted publicly for community evaluation on LMArena.[1][2] The model is the founding member of Microsoft's first-party MAI model family, which by mid-2026 had grown to include speech, image, transcription, reasoning, and coding models.[3][4]
Microsoft introduced MAI-1-preview alongside a speech model, MAI-Voice-1, in an August 28, 2025 post on the Microsoft AI website titled "Two in-house models in support of our mission." Microsoft positioned MAI-1-preview as a consumer-oriented text model "designed to provide powerful capabilities to consumers seeking to benefit from models that specialize in following instructions and providing helpful responses to their everyday queries."[1] The release was framed as an early step, hence the "preview" label, rather than a finished product, and Microsoft began gathering public feedback by entering the model into LMArena, a crowd-sourced platform on which anonymous models are compared head to head and ranked by human votes.[1][2]
Coverage of the launch emphasized the strategic dimension as much as the technical one: MAI-1-preview was Microsoft's clearest public signal that it intended to build frontier-class models of its own, reducing its reliance on OpenAI while keeping that partnership intact.[5][6] At the same time, Microsoft was explicit that it would continue to operate a multi-model strategy, drawing on "the very best models from our team, our partners, and the latest innovations from the open-source community" inside products such as Copilot.[1]
Microsoft AI is the division Microsoft created to house its consumer AI products, including Copilot, Bing, and Edge, with Mustafa Suleyman, a co-founder of DeepMind and Inflection AI, serving as its chief executive after joining Microsoft in March 2024. For years Microsoft's most capable consumer-facing AI features were powered by OpenAI models, the product of a multibillion-dollar partnership and investment. MAI-1-preview, together with MAI-Voice-1, marked the point at which Microsoft began shipping foundation models it had trained itself, end to end.[1][5]
Suleyman has described the effort as a pursuit of "long term self-sufficiency for Microsoft and our partners," and has tied it to Microsoft's ambition to build its own advanced systems. In comments reported in 2026, he said the company "was only sort of set free from our contract with OpenAI about six months ago to formally pursue superintelligence," a reference to a revised agreement between the two companies that loosened earlier contractual restrictions.[6][7] Microsoft has continued to use OpenAI's models, including GPT-5, in its products even as it builds out the MAI stack, and the in-house models are intended to complement partner and open-source systems rather than to replace them outright.[1][6]
Microsoft disclosed relatively little about MAI-1-preview's internals, describing it at a high level as "an in-house mixture-of-experts model" that was "pre-trained and post-trained on ~15,000 NVIDIA H100 GPUs."[1] A mixture-of-experts architecture routes each input through only a subset of the model's parameters, which can lower the computational cost of inference relative to a dense model with the same total parameter count. Microsoft framed the model around efficiency and a consumer focus, stating that the team optimized for instruction-following and useful everyday responses rather than chasing a single benchmark score.[1][2]
Beyond the architecture family and the GPU figure, key specifications were not published at launch. Microsoft did not disclose the model's total or active parameter count, its context window, its training-data composition, or detailed benchmark results, and the model was offered as a preview for public testing rather than as a documented production release.[1][2] The 15,000-H100 figure refers to the cluster Microsoft used for pre-training and post-training; it should be read as Microsoft's own characterization, and the company has separately indicated that it is building out much larger GPU clusters for future training runs.[2][5]
The table below summarizes what Microsoft disclosed about MAI-1-preview and what it did not. Undisclosed items are marked as such rather than estimated.
| Attribute | Detail |
|---|---|
| Developer | Microsoft AI |
| Announced | August 28, 2025 |
| Model type | Text large language model (preview) |
| Architecture | Mixture of experts |
| Total / active parameters | Not disclosed |
| Context window | Not disclosed |
| Training hardware | ~15,000 Nvidia H100 GPUs (pre-training and post-training) |
| Training-data details | Not disclosed |
| Public evaluation | LMArena (debuted at #13 on the Text leaderboard) |
| Access | Public testing on LMArena; limited API by application; planned rollout to select Copilot text tasks |
| Stated focus | Instruction-following, helpful consumer responses, efficiency |
Sources for the table: Microsoft AI announcement and LMArena.[1][2]
Microsoft made MAI-1-preview available for public testing on LMArena at announcement, and in late August 2025 LMArena reported that the model had "broken into the Top 15" of its Text leaderboard, debuting at rank 13. LMArena noted that the Text arena is one of its most competitive categories and characterized a top-15 debut from a new provider as a notable result.[2] Microsoft also opened a limited application-based API to trusted testers and said it would roll MAI-1-preview out for "certain text use cases" within Copilot over the following weeks to learn from real-world usage.[1]
Commentators generally treated the debut as a credible first step rather than a frontier-beating result. A number 13 placement put MAI-1-preview in the mid-pack of the public leaderboard, below the leading frontier systems from OpenAI, Google, and Anthropic, and observers read it as evidence that Microsoft could train a competitive model in-house without claiming that it had yet matched the top of the field.[5][6] Because LMArena is a dynamic, human-voted ranking rather than a fixed academic benchmark, the specific position is best understood as a snapshot from the time of the model's debut.[2]
MAI-1-preview is the founding text model of Microsoft's first-party MAI family, which expanded considerably over the following year. The August 28, 2025 announcement paired it with MAI-Voice-1, a speech generation model that Microsoft said could produce a full minute of audio in under a second on a single GPU and which was integrated into Copilot Daily, Copilot Podcasts, and Copilot Labs.[1] In October 2025 Microsoft added MAI-Image-1, its first in-house text-to-image model, which the company said debuted in the top 10 of LMArena's text-to-image leaderboard.[8]
At its Build 2026 developer conference on June 2, 2026, Microsoft AI's "Superintelligence Team" unveiled a broader family of seven first-party models spanning image, voice, transcription, reasoning, and coding. These included MAI-Thinking-1, described as Microsoft's first reasoning model and a sparse mixture-of-experts system with roughly 35 billion active parameters and a 256,000-token context window; MAI-Code-1 and a faster MAI-Code-1-Flash variant aimed at GitHub Copilot and Visual Studio Code; updated MAI-Voice-2 and MAI-Image-2.5 models; and a multilingual MAI-Transcribe-1.5 speech-to-text model.[3][4] Microsoft reported that MAI-Thinking-1 drew even with Anthropic's Claude Sonnet 4.6 in blind human evaluations run by its independent rating partner and matched Claude Opus 4.6 on a widely cited software-engineering benchmark, though such vendor-reported comparisons should be treated with the usual caution.[3]
MAI-1-preview is significant less for its leaderboard position than for what it represented: the first foundation model Microsoft trained entirely on its own, a tangible step in a strategy of building proprietary frontier models under Mustafa Suleyman's division and reducing structural dependence on OpenAI.[5][6] Its mid-tier LMArena debut showed that Microsoft, despite arriving later than dedicated frontier labs to in-house pretraining, could produce a competitive model, and it established the naming and release cadence for a family that grew quickly to cover speech, images, transcription, reasoning, and code.[2][3]
The model also illustrated the multi-model approach Microsoft has adopted across its products, in which in-house systems such as the MAI family, partner models such as those from OpenAI, and open-source models are orchestrated together rather than relying on any single supplier.[1][6] Within that picture, MAI-1-preview functioned as a proof of concept and a foundation: a preview-stage system whose successors, particularly the reasoning and coding models introduced at Build 2026, carried forward Microsoft's stated goal of long-term AI self-sufficiency.[3][6]