Microsoft’s AI division, called Microsoft AI, has just rolled out its first-ever in-house models: MAI-1-preview for language tasks and MAI-Voice-1 for speech. It might seem like a natural evolution, but it’s also a meaningful shift. Until now, Microsoft’s AI products have leaned heavily on its partnership with OpenAI. That relationship isn’t ending, but clearly, Microsoft is setting the stage for more autonomy.
Key Takeaways
• Microsoft has introduced its first proprietary AI models: MAI-1-preview for text and MAI-Voice-1 for speech.
• MAI-1-preview uses a mixture-of-experts architecture, trained on about 15,000 NVIDIA H100 GPUs, and is available for public testing on LMArena.
• MAI-Voice-1 can generate one minute of high-fidelity audio in under a second using just one GPU.
• Both models are already integrated into Microsoft Copilot, enhancing features like Copilot Daily and Podcasts.
• This marks a strategic shift toward building Microsoft’s own AI capabilities and reducing dependence on OpenAI.
Starting with MAI-1-preview, the company developed this as a mixture-of-experts model, trained across roughly 15,000 Nvidia H100 GPUs. That’s a sizable investment, though still smaller than what some other firms, like xAI with its Grok model, have thrown at training. MAI-1-preview is designed to follow instructions and provide useful, everyday responses. You can actually try it out yourself—the model is being publicly tested on LMArena, a crowdsourced model evaluation platform.
Then there’s MAI-Voice-1, which is arguably just as significant. It’s Microsoft’s first proprietary speech generation model. The company says it can generate a full minute of expressive, high-fidelity audio in less than a second, using only one GPU. That makes it one of the most efficient systems of its kind, at least based on current performance metrics. It’s already powering tools like Copilot Daily and Podcasts, where AI-generated voices narrate the latest news or explain complex topics in a more engaging way. There are also demos available in Copilot Labs, where users can experiment with different tones, storytelling styles, and even character voices.
What’s most notable here is the broader strategy behind these launches. For years, Microsoft’s AI approach has focused on product integration rather than model development. By creating its own models, it gains more control over everything from optimization to deployment, especially for products like Microsoft 365 and Copilot. That’s a major step toward long-term independence.
Mustafa Suleyman, who now leads Microsoft AI, has emphasized that the team used techniques from the open-source community to build these models efficiently. He also mentioned that newer versions are already underway, this time using even more powerful chips like Nvidia’s GB200. It seems they’re preparing for scale, but they’re doing so incrementally.
Still, Microsoft insists it will continue using the best models available, whether those come from its own labs, external partners, or open-source contributors. This isn’t about replacing OpenAI outright, it’s more about having options and being able to steer the ship when needed.
All in all, this is a careful but meaningful move. Microsoft isn’t just investing in AI tools; it’s investing in the infrastructure to build them. And if these first models prove successful, it could change how the company competes in the rapidly evolving AI market.
FAQs
Q. What is the difference between MAI-1-preview and MAI-Voice-1?
A. MAI-1-preview is a text-based model designed for handling queries and instructions, while MAI-Voice-1 is a speech generation model that creates expressive audio.
Q. Are these models part of Microsoft Copilot?
A. Yes, Microsoft has started integrating both models into its Copilot AI assistant for various text and voice-based functions.
Q. Where can I test the new Microsoft AI models?
A. You can test MAI-1-preview on the LMArena platform, which is a public community testing ground for large language models. MAI-Voice-1 can be tried through Copilot Labs.
Q. Why is Microsoft creating its own AI models now?
A. Microsoft is creating its own models to have more control over its AI technology, improve integration with its products, and reduce its dependence on third-party partners like OpenAI.
