MAI-Thinking-1

MAI-Thinking-1 is a reasoning model developed by Microsoft AI, unveiled at Microsoft Build 2026 on June 2, 2026 as the company's first in-house flagship reasoning system. ^[1][2] Built by Microsoft's "MAI Superintelligence Team" under Mustafa Suleyman, the model is a 35-billion-active-parameter mixture-of-experts network with roughly one trillion total parameters and a 256,000-token context window. ^[1][3][11] Microsoft says it was trained from scratch on commercially licensed data with no distillation from other companies' models, a deliberate contrast with the way many rivals bootstrap new systems. ^[1][2] At launch it was offered in private preview on Microsoft Foundry; on August 12, 2026 Microsoft moved it to public preview in the Foundry model catalog, priced at $2 per million input tokens and $8 per million output tokens. ^[1][12][15] Microsoft positioned the model as central to a broader push to reduce its dependence on OpenAI. ^[2][4]

Overview

MAI-Thinking-1 was announced as the headline entry in a family of seven new first-party models released at Build 2026, alongside coding, image, voice, and transcription systems. ^[2][4] It is the first Microsoft model explicitly built for step-by-step chain-of-thought reasoning rather than the instruction-following and consumer Copilot scenarios that defined the company's earlier in-house releases. ^[1][5] Microsoft describes the target use cases as complex multi-step instructions, long-context reasoning, mathematical problem solving, and code generation. ^[1][3] The Foundry catalog listing adds that the model allocates reasoning effort "adaptively according to prompt complexity" and that chain-of-thought sequences used in post-training were "length-compressed" to reduce the token cost of a typical response. ^[12]

The model fits into a strategy Suleyman, the CEO of Microsoft AI, summarized at the event as "long term self-sufficiency for Microsoft and our partners." ^[4] For most of the past decade Microsoft built its AI products on top of OpenAI's models, later adding Anthropic as a second supplier. MAI-Thinking-1 is the clearest sign yet that the company intends to supply at least part of that demand from its own labs. ^[2][4] Microsoft renegotiated its OpenAI partnership in April 2026, ending Microsoft's exclusive license to OpenAI intellectual property, which removed one barrier to shipping competing first-party frontier models. ^[18][6]

The MAI team and lineage

MAI-Thinking-1 comes from the same Microsoft AI group that began publishing in-house models in 2025. In August 2025 the team released its first two: MAI-Voice-1, an efficient speech-generation model used in features such as Copilot Daily, and MAI-1-preview, a text foundation model aimed at instruction following and consumer Copilot interactions. ^[7][8] MAI-1-preview was trained on roughly 15,000 Nvidia H100 GPUs, a smaller training footprint than the headline figures cited by some competitors, and it was submitted for public benchmarking on platforms such as LMArena. ^[7]

The group was reorganized around an explicit goal of "humanist superintelligence," and the unit behind the Build 2026 releases is referred to as the MAI Superintelligence Team. ^[1][5] Suleyman, who co-founded DeepMind and later Inflection AI before joining Microsoft in 2024 to lead its consumer AI work, framed the team's output as state-of-the-art capability "explicitly designed to serve people and organizations, and not to replace them." ^[5][9] MAI-Thinking-1 represents the team's move up the capability ladder, from a voice model and a mid-tier consumer text model to a frontier-class reasoning system.

Architecture and specifications

Microsoft describes MAI-Thinking-1 as a sparse mixture-of-experts model with 35 billion active parameters and approximately one trillion total parameters, meaning only a fraction of the network is engaged for any given token. ^[1][3] That design gives it a smaller inference footprint than dense models of comparable quality, which underpins Microsoft's repeated emphasis on cost efficiency. ^[3][10] The context window is 256,000 tokens, enough to process roughly a 600-page document in a single pass. ^[1] Every Microsoft source gives the same figure: the June 2 announcement ("a 256k token window"), Suleyman's Build keynote ("a 35B active parameter MoE with a 256K context window"), the technical report ("after mid-training, MAI-Thinking-1 achieves a maximum context length of 256K"), the Foundry catalog (256,000-token context, 64,000-token output limit) and the Microsoft Learn model table. ^{[1][5][11][12][13]} The 128K figures in the technical report are training and evaluation settings, not the model's window: the GraphWalks long-context benchmark is scored on its 128K-and-under subset, the reinforcement-learning stage capped generations at 128K output tokens, and an appendix table (Table 17) compares scores at 128K versus 256K maximum output tokens. ^[11]

The technical report, "MAI-Thinking-1: Building a Hill-Climbing Machine," describes the underlying base model, which Microsoft calls MAI-Base-1. ^[11] It is a decoder-only transformer that pairs five local-attention layers (sliding window of 512 tokens, with rotary position encoding) with one global-attention layer that uses no position encoding at all, a layout the report credits to Gemma 3. Attention is grouped-query attention with 8 key-value heads and a per-head dimension of 128. Feed-forward blocks alternate between dense layers and high-sparsity MoE layers in which 8 of 512 experts are activated per token in a compressed latent space (the "LatentMoE" design). Input and output embeddings are tied, and the tokenizer is OpenAI's o200k_base with a 200,019-entry vocabulary, chosen "to simplify integration with existing in-house tools." ^[11] The report says pre-training ran for 30 trillion tokens, followed by mid-training phases totaling 3.55 trillion tokens that emphasized STEM, mathematics and code and extended the context length to 256K. ^[11]

Sources: Microsoft AI announcement and model page, technical report, Foundry catalog, Microsoft Learn and the Foundry blog. ^{[1][3][11][12][13][15]}

On the developer side the model is compatible with the widely used Chat Completions API, and it supports function calling and multi-layered developer instructions. ^[1][3] The Foundry catalog lists the model as primarily optimized for English, with support for German, Spanish, French, Italian, Portuguese, Simplified Chinese, Russian, Hindi, Japanese, Korean, Arabic and Hungarian, and "limited coverage" of several others. ^[12] Microsoft also said that, for the first time, developers would be able to tune the weights of one of its in-house models themselves, a capability aimed at enterprises that want a customized system without handing their workflows to a third party. ^[3][10]

Training approach

The most heavily promoted aspect of MAI-Thinking-1 is how it was trained. Microsoft's June 2 announcement said the model was built "from the ground up on enterprise grade, clean and commercially licensed data, without distillation from third-party models," and that AI-generated content was excluded from pre-training. ^[16] Microsoft later reworded the post (a June 8 revision, still carried on the live page) to "We don't distill from other labs and we don't rely on opaque data. Our datasets are clean, traceable, and enterprise-grade." ^[1][16] The technical report is more specific: pre-training used "publicly available and licensed human-generated data covering web data, public GitHub code, books, academic papers, news, multilingual text, and domain-specific materials," all processed in-house, with no synthetic data generated by language models and no open-source training datasets. ^[11] Suleyman stressed that there was no distillation from other companies' models, including OpenAI's GPT series, a pitch aimed squarely at enterprises that care about clean data lineage and the legal provenance of the systems they deploy. ^[2][4]

The reinforcement-learning stage started "from scratch, learning to reason with no prior exposure to reasoning traces," according to the report, and produced three domain specialists (STEM reasoning, agentic coding and tool use, and helpfulness and safety) that were then consolidated. ^[11] Microsoft frames the whole pipeline as a "hill-climbing machine," a repeatable loop of data, rewards, environments and evaluations meant to keep improving capabilities over time. ^[1][11]

A public data summary, published under the EU AI Act template and dated August 12, 2026, adds detail on the data. It names Microsoft Ireland Operations Limited as the provider, lists GitHub public repositories, Wikipedia and CommonCrawl among the public datasets, puts the text training corpus in the more-than-10-trillion-token band, gives March 2026 as the date the training dataset was first used, and says datasets were "collected as late as July 2026 for model training." ^[14] That last date, echoed by a "July 2026" training cut-off in the Foundry catalog, postdates the June 2 announcement; Microsoft's published materials do not address the discrepancy; the summary says synthetic data is used in post-training and in pre-training data curation (classification, captioning, metadata), while the technical report says no language-model-generated synthetic data was used as pre-training tokens. ^[12][14]

Microsoft framed this as a question of trust and control rather than only quality. In the keynote the company argued that customers who train and tune on MAI models keep the benefits of their own workflows and control the resulting model, in contrast with building on a provider whose interests may diverge from theirs. ^[5] Microsoft also said the model "climbed entirely from the bottom, without specifically targeting any of these benchmarks, and with zero distillation," a claim meant to distinguish genuine capability from benchmark optimization. ^[5] These are Microsoft's characterizations; as of the announcement the company had published technical materials, including a model card and a paper, but the results had not been independently reproduced by outside labs. ^[1][2]

Benchmarks

Microsoft reported strong results on mathematics and competitive coding evaluations relative to the model's size. On the American Invitational Mathematics Examination (AIME), MAI-Thinking-1 scored 97.0 percent on AIME 2025 and 94.5 percent on AIME 2026. ^[1][3] On SWE-Bench Pro, a software-engineering benchmark, Microsoft reported 52.8 percent (rounded to "53%" in the keynote) and said the model matches Anthropic's Claude Opus 4.6 on coding tasks. ^[1][5][11] In blind side-by-side human evaluations run by Surge, an independent rating partner, raters preferred MAI-Thinking-1 over Claude Sonnet 4.6 for overall quality across single-turn and multi-turn tasks. ^[1][2]

The technical report gives the fuller picture. Its main results table lists MAI-Thinking-1 alongside competitor numbers "taken from respective official model cards," and the report's own summary is that the model "does not lead the field, but it delivers consistently strong performance across a broad set of benchmark categories." ^[11] All MAI-Thinking-1 scores are averages over four runs at temperature 1 and top-p 0.97. ^[11]

Sources: Microsoft AI technical report (Tables 11 and 14), announcement, model page, and Build 2026 keynote. Competitor figures are as quoted by Microsoft from those models' own cards. ^{[1][3][5][11]}

Two points in the report qualify the launch messaging. First, the same table shows GPT 5.4 (57.7 percent), Kimi K2.6 (58.6 percent), DeepSeek V4 (55.4 percent) and GLM-5.1 (58.4 percent) ahead of MAI-Thinking-1 on SWE-Bench Pro, so "toe-to-toe with Claude Opus 4.6" is a comparison with one model rather than a leaderboard claim. ^[11] Second, while raters preferred MAI-Thinking-1 to Sonnet 4.6 (overall preference +0.07 on a scale from -1.5 to 1.5), they preferred Opus 4.6 to MAI-Thinking-1 by the same margin; the report attributes MAI-Thinking-1's wins mainly to conciseness, relevance, style and tone, with instruction following, factuality and completeness "roughly equivalent (within noise)." ^[11] On Microsoft's own evaluation suite the model trails Sonnet 4.6 on long-context retrieval (GraphWalks up to 128K: 90 vs 96), tool calling (BFCL v3: 72 vs 76) and the health benchmarks, while leading on IFBench (69 vs 50) and SimpleQA Verified (31 vs 29). ^[11]

Microsoft presented the human-preference result as evidence that benchmark scores were translating into practical usefulness rather than narrow test performance. ^[2] The company also pointed to a customer-specific result: after tuning MAI models for the consulting firm McKinsey's tasks, Suleyman said they outperformed OpenAI's GPT-5.5 on quality with roughly ten times lower cost, a projection the keynote footnotes as based on public GPT pricing and MAI pricing data "scaled across model sizes." ^[4][5] As with the other figures, these are vendor-reported numbers and should be read with the usual caution that applies to launch-day benchmarks.

Strategic significance

The release matters less for any single benchmark than for what it signals about Microsoft's direction. Microsoft has spent years and many billions of dollars building products on OpenAI's models, and more recently it added Anthropic models, including Claude Opus 4.8, to its Foundry catalog. ^[2] MAI-Thinking-1 shows Microsoft trying to build a credible in-house alternative so it can rely less on suppliers whose commercial interests do not always align with its own. ^[2][4]

Cost is the other half of the pitch. Microsoft repeatedly described MAI-Thinking-1 as a high-efficiency, low-token-cost model and positioned it as the most cost-efficient frontier-class option in its tier, an argument aimed at budget-conscious enterprise buyers weighing per-token bills across providers. ^[3][10] The company tied the efficiency story to its own silicon, citing co-design with its Maia 200 accelerator as part of a full-stack ownership strategy spanning chips, models, and tools; Suleyman said the team was benchmarking the model on Maia 200 "head-to-head against the GB200" and seeing a further 1.4x performance-per-watt gain. ^[5] If the model holds up under independent testing, it gives Microsoft leverage in pricing negotiations and a fallback should its relationships with frontier labs sour.

There is real skepticism to weigh against the launch claims. MAI-Thinking-1 is Microsoft's first reasoning model, its benchmark results have not been independently verified, and a single flagship release does not by itself replace the breadth of capabilities Microsoft sources from OpenAI and Anthropic. What is clear is the intent: Microsoft wants to own more of its AI stack, and a frontier reasoning model trained without anyone else's data is the most direct statement of that ambition it has made so far.

Availability

At announcement, MAI-Thinking-1 was available in private preview on Microsoft Foundry, the company's platform for integrating models into applications, with access by invitation. ^[1][3][16] Microsoft also said the model would be accessible through a MAI Playground for public preview, and that its MAI models would be distributed through third-party platforms including Fireworks AI, Baseten, and OpenRouter. ^[3][5] Microsoft did not publish standard per-token pricing for MAI-Thinking-1 at launch, though it emphasized low token cost relative to comparable frontier models; the Foundry launch post's early revisions listed the model only as "In private preview, request access here." ^[3][10][17]

On August 12, 2026 Microsoft updated the announcement with a banner reading "MAI-Thinking-1 is now available in public preview. Try it now in Microsoft Foundry," and the Foundry catalog lists August 12, 2026 as the model's release date. ^[1][12] The same day the Foundry blog post was revised to say the model is "available directly in the Foundry Model Catalog" with pricing "starting at $2 USD per 1M tokens for input and $8 USD per 1M tokens for output," and Microsoft's EU data summary gives August 12, 2026 as the date the model was placed on the EU market. ^[14][15] The Microsoft Learn model table lists it as a preview chat-completion model with reasoning content, tool calling, text output up to 64,000 tokens and a 256,000-token context length. ^[13] The model is also served in the MAI Playground under the identifier mai-thinking-1-latest. ^[3]

References

1. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17 ^18 ^19 ^20 ^21 ^22 ^23 ^24Introducing MAI-Thinking-1. Microsoft AI, June 2, 2026 (updated August 12, 2026).
2. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11Microsoft Build 2026: MAI-Thinking-1 Is First In-House Reasoning Model, Trained Without OpenAI Data. Tech Times, June 2, 2026.
3. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14MAI-Thinking-1. Microsoft AI model page.
4. ^1 ^2 ^3 ^4 ^5 ^6 ^7Microsoft unveils seven homegrown AI models in new bid for 'long term self-sufficiency'. GeekWire, June 2, 2026.
5. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12Microsoft Build 2026: MAI Keynote Transcript. Microsoft AI, June 2, 2026.
6. ^Microsoft unveils new AI models to lessen reliance on OpenAI and lower costs for developers. CNBC, June 2, 2026.
7. ^1 ^2Microsoft reveals two in-house AI models: MAI-Voice-1 and MAI-1-preview. Neowin, August 28, 2025.
8. ^Two new in-house models in support of our mission. Microsoft AI, August 28, 2025.
9. ^Microsoft unveils MAI-Thinking-1 reasoning and MAI-Code-1 coding models. Neowin, June 2, 2026.
10. ^1 ^2 ^3 ^4Microsoft Expands Foundry With Seven In-House MAI Models. WinBuzzer, June 2, 2026.
11. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17 ^18MAI-Thinking-1: Building a Hill-Climbing Machine (technical report, PDF). Microsoft AI (The Microsoft AI Team), June 2026 (published June 2, 2026; revised, current file dated June 6, 2026).
12. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8MAI-Thinking-1 | Model Catalog | Microsoft Foundry. Microsoft Foundry model catalog, model version 2026-06-01, accessed September 6, 2026.
13. ^1 ^2 ^3Microsoft models sold by Azure. Microsoft Learn, September 4, 2026.
14. ^1 ^2 ^3 ^4Data Summary for MAI-Thinking-1 (PDF). Microsoft, version 1.0, August 12, 2026.
15. ^1 ^2 ^3 ^4New MAI models in Microsoft Foundry across text, image, voice and speech. Microsoft Tech Community (Azure AI Foundry Blog), June 2, 2026, updated August 12, 2026 (version 5.0).
16. ^1 ^2 ^3 ^4 ^5Introducing MAI-Thinking-1 (June 2, 2026 version). Internet Archive capture of the Microsoft AI post, June 2, 2026.
17. ^New MAI models in Microsoft Foundry across text, image, voice and speech (June 8, 2026 version). Internet Archive capture of the Microsoft Tech Community post, version 2.0, June 8, 2026.
18. ^The next phase of the Microsoft-OpenAI partnership. Microsoft (Official Microsoft Blog), April 27, 2026.