Microsoft unveils first in-house reasoning model: MAI-Thinking-1 with one trillion parameters trained without OpenAI data, turns Windows into AI agent operating system
What it really says
At the Build 2026 developer conference on June 2-3, Microsoft unveiled seven new in-house AI models, led by MAI-Thinking-1, the company's first reasoning model. MAI-Thinking-1 is based on a sparse mixture-of-experts architecture with 35 billion active parameters out of approximately one trillion total parameters and a 256,000-token context window. The model was trained entirely on commercially licensed data without distillation or training data from third-party models such as OpenAI's GPT series. On the demanding coding benchmark SWE-Bench Pro, MAI-Thinking-1 scores 53 percent, matching Anthropic's Claude Opus 4.6. On the mathematics benchmark AIME 2025, it reaches 97 percent, and 94.5 percent on AIME 2026. In 1,350 blind side-by-side evaluations by professional raters from Surge, MAI-Thinking-1 was preferred over Claude Sonnet 4.6. Alongside the reasoning model, Microsoft introduced six additional models: MAI-Image-2.5 for image generation, MAI-Voice-2 for speech synthesis in over 15 languages, MAI-Transcribe-1.5 for transcription in 43 languages, MAI-Code-1 as a specialized coding model for GitHub Copilot, and Flash variants for image and voice. Additionally, Microsoft announced the Windows Agent Runtime, a free update for Windows 11 Pro arriving in September 2026 that transforms Windows into a runtime environment for AI agents. The Agent Runtime provides APIs and services that allow applications to host local agents capable of interacting with the operating system, other apps, and the cloud.
Our assessment
This announcement deserves attention from two perspectives. First, it marks a strategic turning point: Microsoft, which invested $13 billion in OpenAI, is now building its own model family that operates independently of OpenAI technology. This is reassuring in one sense because it strengthens competition in the AI market and reduces dependency on a single provider. On the other hand, it demonstrates how rapidly AI development is advancing: a model with one trillion parameters that would have been science fiction two years ago is now just one of seven products unveiled simultaneously. Second, the Windows Agent Runtime represents a fundamental paradigm shift: Windows evolves from a passive operating system to an active platform for AI agents that autonomously interact with apps, files, and cloud services. This raises legitimate questions: Who controls what an AI agent does on your computer? What data does it send to the cloud? How do you prevent a malfunctioning agent from causing damage? The benchmark results are impressive but have not yet been independently reproduced. Microsoft published a preprint describing the evaluation methodology, but external confirmation is pending. The claim that MAI-Thinking-1 matches Opus 4.6 should therefore be viewed with caution.
Relevance for Germany
For Germany, this announcement has several implications. First, Microsoft with Windows, Office 365, and Azure is deeply embedded in German enterprise IT. When Windows itself becomes an AI agent platform, this affects millions of workplaces in German offices. Companies must ask whether they are ready to grant AI agents access to their business processes and what compliance requirements this entails under the EU AI Act and GDPR. Second, the new MAI models offer an alternative to OpenAI, which is relevant for German companies seeking to diversify their dependency on a single AI provider. The models' availability on OpenRouter, Fireworks, and Baseten expands the choices. Third, the Windows Agent Runtime sits in tension with the EU's newly proposed Cloud and AI Development Act: if AI agents run locally on Windows but send data to Microsoft servers, the sovereignty question arises anew. The Bundesnetzagentur as Germany's future AI supervisory authority will need to address whether and how agent-based AI systems fall within the scope of the EU AI Act. Fourth, the announcement puts pressure on the European AI industry: while Microsoft introduces seven models simultaneously, Europe has no comparable model in this performance class.
Fact check
The primary source is Microsoft's official announcement on microsoft.ai and the Microsoft blog. The technical specifications of MAI-Thinking-1, including 35 billion active parameters, approximately one trillion total parameters, sparse MoE architecture, and 256,000-token context window, are consistently confirmed by CNBC, TechTimes, Thurrott, and the Latent Space analysis. The benchmark results of 53 percent on SWE-Bench Pro and 97 percent on AIME 2025 come from Microsoft's own preprint and have not yet been independently reproduced. The information that the model was trained without OpenAI data is confirmed by CNBC and TechTimes citing Microsoft. The Windows Agent Runtime announcement for September 2026 comes from the Build 2026 keynote and is confirmed by windowsnews.ai and directionsonmicrosoft.com. Availability on platforms like OpenRouter, Fireworks, and Baseten is mentioned in multiple independent reports.
Source
- • https://www.cnbc.com/2026/06/02/microsoft-unveils-new-ai-models-lessen-reliance-on-openai-lower-costs.html
- • http://www.techtimes.com/articles/317631/20260602/microsoft-build-2026-mai-thinking-1-first-house-reasoning-model-trained-without-openai-data.htm
- • https://microsoft.ai/news/introducing-mai-thinking-1/
- • https://www.heise.de/en/news/Microsoft-Build-2026-AI-development-with-under-and-for-Windows-11315701.html