Microsoft introduces its first internal AI models for Copilot to compete with GPT-5

Microsoft’s Artificial Intelligence arm has unveiled its first in-house models, marking a significant step in the company’s push to expand its AI offerings beyond those developed by OpenAI. The new releases, called MAI-Voice-1 and MAI-1-preview, were announced on Thursday and are already beginning to show up across Microsoft’s products.

The MAI-Voice-1 variant has remarkable speed and is intended for speech generation. Microsoft claims that it can use just one GPU to create a whole minute of audio in less than a second. Because of this, it is significantly more effective than the majority of speech AI systems on the market today. The technology has already been used by the business. It powers Copilot Daily, which delivers top news stories with an AI narrator, as well as tools that generate podcast-style discussions to break down complex subjects.

MAI-Voice-1 is available for experimentation on Copilot Labs, Microsoft’s testing ground for new AI features, for anyone who is interested. In addition to selecting what they want the AI to say, users can also alter the generated voice’s tone and style by typing in a prompt.

A larger text-based model called MAI-1-preview was also released by Microsoft at the same time. It was trained on an astounding 15,000 Nvidia H100 GPUs. The goal of this system is to manage tasks that require following instructions and provide conversational answers to common inquiries. The business refers to it as “a glimpse of future offerings inside Copilot,” implying that in the upcoming months, the model will play a larger part across all of its AI assistant services.

Microsoft looks determined to bring more of the work in-house, even if it has relied largely on OpenAI’s massive language models to run Copilot. For several text-based use cases, the MAI-1-preview paradigm has already begun to be implemented in limited form. It is also undergoing public evaluation through the AI benchmarking platform LMArena, where developers and researchers can test its performance.

The action highlights Microsoft’s consumer-first AI strategy. The company’s AI leader, Mustafa Suleyman, described this approach in an interview with the Decoder podcast last year. According to Suleyman, “we have to make something that is really optimized for our use case and works incredibly well for the consumer.” Therefore, we have a ton of extremely valuable and predictive data on the ad side, consumer telemetry, and other areas. My focus is on building models that really work for the consumer companion.”

This emphasis on personal use cases means the models are not being developed primarily with enterprise customers in mind. Rather, Microsoft is presenting them as virtual friends who can explain complex subjects, narrate news, and serve as conversational helpers in daily life.

Microsoft emphasized its long-term goals in a blog post unveiling the models, saying, “We have big ambitions for where we go next.” In addition to pursuing additional developments in this area, we think that coordinating a variety of specialized models to meet various user intents and use cases will yield enormous benefits.

Microsoft is indicating with these initial actions that it wants Copilot to serve as a showcase for its own AI research and become less reliant on OpenAI. As more individuals test and utilize these models, it will become increasingly evident if they meet the aspirations.