Microsoft AI team develops AI inference model MAI, performing on par with OpenAI

According to sources familiar with the matter, a team led by Kalkaman Simonyan, the deputy of Microsoft's (MSFT. O) AI leader Suleiman, recently completed training on a series of Microsoft models known internally as MAI, which perform almost as well as OpenAI and Anthropic's main models in recognized benchmark tests. The team is still training an inference model that uses thought chain technology to think about problems before solving them, which can directly compete with OpenAI. This insider said that the team has started trying to replace OpenAI's model with the MAI model (which is much larger in scale than early Microsoft Phi models) in Microsoft's Copilot. This plan will enable Microsoft's models to directly compete with similar API products from OpenAI and other artificial intelligence labs. In addition, according to informed sources, Microsoft has tested alternative models for Copilot, including models from Anthropic and Musk's xAI, as well as open-source models from DeepSeek and Meta platforms.