Introduction
MAI-1-preview is Microsoft’s in-house AI model now in public testing; it aims to reduce reliance on OpenAI and potentially strengthe
Context
Historically, Microsoft has used OpenAI models for features in Bing and Windows 11; MAI-1-preview represents an effort to diversify and control core AI technology while maintaining strategic ties with OpenAI.
Why it matters
Building an internal model may lower commercial dependency, speed product integration, and provide strategic alternatives, but it raises questions around performance and scale.
Technical details and compute
Microsoft refined MAI-1-preview using roughly 15,000 Nvidia H100 GPUs and operates a cluster with Nvidia GB200 chips. On LMArena the model currently ranks 13th for text workloads.
The challenge
The main challenge is scaling an in-house model to compete with established offerings from OpenAI, Anthropic, Google and others, while controlling costs and maintaining product integration.
Approach
Microsoft is running public tests on LMArena and accepting developer early-access requests, while continuing investments in compute and model improvements.
"We have big ambitions for where we go next — model advancements, an exciting roadmap of compute, and the chance to reach billions of people through Microsoft’s products."
Mustafa Suleyman, CEO of Microsoft’s AI unit
Product and market impact
A mature MAI-1-preview could give Microsoft more control over assistant capabilities, privacy, cost structure and feature roadmaps, while intensifying competition with OpenAI.
- Integration potential in Copilot for specific text use cases
- Competitive dynamics with OpenAI and other AI vendors
- Heavy reliance on high‑performance GPU infrastructure
Limitations and risks
Current public rankings and tests show room for improvement; balancing partnership commitments with OpenAI and pursuing internal models adds strategic complexity.
Conclusion
MAI-1-preview is a strategic move by Microsoft to build internal AI capability; public testing and developer access will shape its evolution and product adoption.
FAQ
Short answer: MAI-1-preview is Microsoft’s first foundation model trained entirely in-house and is available for public evaluation.
- How can developers request access to MAI-1-preview? Microsoft published a form for early access requests; check the company blog for the latest details.
- Will MAI-1-preview be used in Copilot? Microsoft said it will roll MAI-1-preview out for certain text use cases within Copilot in the coming weeks.
- What compute was used to train MAI-1-preview? The model was refined with around 15,000 Nvidia H100 GPUs and also uses a GB200 cluster.
- Does MAI-1-preview reduce Microsoft’s reliance on OpenAI? It could reduce dependency if scaled successfully, but Microsoft remains a major investor and partner of OpenAI.