Microsoft has announced the launch of two of its own artificial intelligence models that were entirely built and trained by its internal teams. reports Engadget.
The first of them, MAI-Voice-1, is a model for generating natural speech. The second — MAI-1-preview — is a text model that became Microsoft’s first fundamental AI solution developed from scratch. MAI-Voice-1 is already integrated into the Copilot Daily and Podcast features. The MAI-1-preview text model is currently available for testing on the LMArena platform and is planned to be used in selected Copilot scenarios in the future.
The head of the Microsoft AI division, Mustafa Suleiman, noted that both models were created with an emphasis on efficiency and economy. So, MAI-Voice-1 runs on a single GPU, while MAI-1-preview was trained on about 15,000 Nvidia H-100 GPUs. For comparison, the Grok model from xAI needed more than 100 thousand such chips.
According to Suleiman, the key in the process of training models was the careful selection of data that really contribute to performance, without unnecessary expenditure of computing resources on inefficient tokens.
Although Microsoft Copilot is still based on GPT technologies from OpenAI, the launch of its own models demonstrates the company’s desire for greater autonomy in the field of artificial intelligence. Suleiman also said that Microsoft is implementing a five-year AI development plan, which is funded on a quarterly basis.
