Microsoft has taken a big step in its AI journey by launching two new models developed entirely in-house. After years relying on OpenAI’s foundation models, the tech giant now aims to create its own AI tools for consumer apps. This move shows Microsoft’s desire to have more control over its AI technology.

Introducing MAI-Voice-1 and MAI-1-preview

Microsoft announced MAI-Voice-1 and MAI-1-preview this week. MAI-Voice-1 is a speech generation model that can produce a minute of audio in less than a second. It runs efficiently on a single GPU and is already being used in Microsoft’s Copilot Daily to generate news summaries. It’s also part of the Copilot Podcasts feature, which lets users create podcasts instantly from prompts.

In a demonstration, Microsoft showed that MAI-Voice-1 can generate realistic, expressive audio. It can handle both single and multiple speakers, making it versatile for different applications. Meanwhile, the MAI-1-preview is currently being tested on LMArena, a community site where users compare AI models. Developers can also apply to access the API for this model, which Microsoft plans to roll out gradually within Copilot.

The Shift Toward In-House AI Development

Microsoft built MAI-1-preview using an in-house mixture-of-experts model. It was pre-trained and refined on 15,000 Nvidia H100 GPUs, a smaller setup compared to some rivals that reportedly used 100,000 GPUs. The company is now running the model on Nvidia’s more powerful GB200 cluster to boost performance.

This move is significant because, until now, Microsoft mainly depended on OpenAI’s models, which it has heavily invested in—around $13 billion since 2019. OpenAI’s models are hosted on Microsoft’s Azure cloud. But in 2024, Microsoft surprised many by hiring Mustafa Suleyman, the founder of DeepMind’s partner Inflection AI, along with his team. This team is working on developing independent AI models for Microsoft, separate from OpenAI.

Microsoft remains tight-lipped about the exact relationship with OpenAI, but Suleyman emphasizes that Microsoft plans to keep options open. He explained that the company might use models from different sources, including third-party developers, open source options, and its own. He envisions an “orchestrator,” a platform that routes queries to the most suitable AI model based on what’s needed. This approach allows Microsoft to combine various models to serve different user needs better.

In the short term, Microsoft is focusing on developing AI for its consumer products, especially within Copilot. The company sees this as a way to unlock new value by managing a range of specialized models tailored for specific tasks. This renewed focus on in-house development signals that Microsoft is serious about building its own AI capabilities, even as it continues to collaborate with OpenAI.

Overall, this move marks a key point in Microsoft’s AI strategy. It shows a clear desire to diversify its AI toolkit and reduce reliance on external models. The company’s investments and new models indicate it’s gearing up for a future where it can control more of its AI technology and better serve its users with tailored solutions.

Inspired by

• https://www.infoworld.com/article/4048487/microsofts-signals-shift-from-openai-with-launch-of-first-in-house-ai-models-for-copilot.html

Sources

• gmpg.org
• microsoft.ai
• microsoft.com
• microsoft.com
• microsoft.com