Microsoft's AI division announced its first in-house model, marking a key step in the company's push to expand its AI offering beyond what Openai developed. The new release, called Mai-Voice-1 and Mai-1-Preview, was announced on Thursday and is already beginning to appear across Microsoft's products.
The MAI-Voice-1 model is designed for voice generation and boasts impressive speeds. According to Microsoft, you can generate one minute of audio in under a second using only one GPU. This makes it much more efficient than most voice AI systems currently available. The company has already made its technology work. Copilot provides Powerthed Daily with top news stories along with AI narrators and tools to generate podcast-style discussions to break down complex subjects.
Anyone who wants to try it can try Mai-Voice-1 at Copilot Labs, Microsoft's testing ground for new AI features. Users can also enter prompts, select what to say to the AI, and customize the tone and style of the generated audio.
In addition to this, Microsoft has introduced MAI-1-PREVIEW. This is a larger text-based model trained on 15,000 incredible NVIDIA H100 GPUs. The system aims to handle instructional tasks and answer everyday questions in a conversational way. The company describes it as a “glimpse into the future offering within Copilot,” suggesting that the model is set to play a bigger role across AI assistant services over the coming months.
Microsoft is leaning heavily towards Openai's large-scale language model and leaning on Power Copilot, but it is now aimed at bringing more work in-house. The MAI-1-PREVIEW model is already beginning to develop in a limited format for certain text-based use cases. It is also publicly recognized through Lmarena AI benchmark platform that allows developers and researchers to test their performance.
The move expands the scope of Microsoft's consumer-first approach to AI. The company's AI chief Mustafa Suleyman explained the strategy in an interview with podcast decoder last year. “My logic is that you have to create something that works very well for the consumer and that is very optimised for your use case,” Suleyman says. “So there's a huge amount of very predictable and very useful data, like advertising sides, consumer telemetry, and so on. My focus is on building models that actually work for the peers of consumers.”
The emphasis on personal use cases means that the model is not developed primarily with enterprise customers in mind. Instead, Microsoft positions them as digital companions who can narrate news, explain tricky topics, and act as conversation assistants in their daily lives.
In a blog post to announce the model, Microsoft highlighted its long-term ambitions. “There is a great ambition where we go next. We believe that not only will we continue to make further progress here, but we will also be able to tweak the different specialist models that will help with the intentions and use cases of different users, which will result in a significant loss of value.”
These first steps indicate that Microsoft hopes that Copilot will no longer depend on Openai and will become a showcase for its own AI research. Whether these models will reach ambitions will become more clear as more people start testing and using them.
– end
