Microsoft Launches Its First In-house Ai Models

Microsoft announced its first homegrown AI models connected Thursday: MAI-Voice-1 AI and MAI-1-preview. The institution says its caller MAI-Voice-1 reside exemplary tin make a minute’s worthy of audio successful nether 1 2nd connected conscionable 1 GPU, while MAI-1-preview “offers a glimpse of early offerings wrong Copilot.”

Microsoft already uses MA1-Voice-1 to powerfulness a mates of its features, including Copilot Daily, which has an AI big singing nan day’s apical news stories, and to generate podcast-style discussions to thief explicate topics. 

You tin effort MA1-Voice-1 retired for yourself on Copilot Labs, wherever you tin participate what you want nan AI exemplary to say, arsenic good arsenic alteration its sound and style of speaking. In summation to this model, Microsoft introduced MAI-1-preview, which it says it trained connected astir 15,000 Nvidia H100 GPUs. It’s built for users successful request of an AI exemplary tin of pursuing instructions and “providing adjuvant responses to mundane queries.” 

Microsoft AI main Mustafa Suleyman said during an section of Decoder past twelvemonth that nan company’s soul AI models aren’t focused connected endeavor usage cases. “My logic is that we person to create thing that useful highly good for nan user and really optimize for our usage case,” Suleyman said. “So, we person immense amounts of very predictive and very useful information connected nan advertisement side, connected user telemetry, and truthful on. My attraction is connected building models that really activity for nan user companion.”

The institution plans connected rolling retired MAI-1-preview for definite matter usage cases successful its Copilot AI assistant, which presently relies connected OpenAI’s ample connection models. It has besides started publically testing its MAI-1-preview exemplary connected nan AI benchmarking level LMArena.

“We person large ambitions for wherever we spell next,” Microsoft writes successful nan blog post. “Not only will we prosecute further advances here, but we judge that orchestrating a scope of specialized models serving different personification intents and usage cases will unlock immense value.”