Microsoft is developing its own artificial intelligence reasoning models to reduce its reliance on OpenAI, the company that has been its partner in the field. The tech giant may also offer these models to outside developers, according to a report released Friday by The Information, citing an unnamed source familiar with the project.
The Redmond, Washington-based company has been looking to reduce its dependence on OpenAI, even though the initial partnership put Microsoft in a leadership position among Big Tech peers in the lucrative AI race. The Information’s source indicated that Microsoft has begun testing models from xAI, Meta, and DeepSeek as potential replacements for OpenAI within its Copilot product.
Reuters previously reported that Microsoft has been working to incorporate internal and third-party AI models to power Microsoft 365 Copilot, with the goal of diversifying its sources and reducing costs. When Microsoft announced 365 Copilot in 2023, a key selling point was that it used OpenAI’s GPT-4 model.
The Information’s report asserts that Mustafa Suleyman, who leads Microsoft’s AI division, has overseen the completion of training for a family of models, internally referred to as MAI. These MAI models are reportedly performing on par with leading models from OpenAI and Anthropic on standard benchmarks. Suleyman’s team is additionally training reasoning models, which utilize chain-of-thought techniques – a reasoning process that involves generating answers by using intermediary steps – that could directly compete with OpenAI’s offerings, the source said.
The report further states that Suleyman’s team is already experimenting with using the MAI models in Copilot, which are significantly larger than the previous Phi models developed by Microsoft. The company is considering releasing the MAI models by the end of the year as an application programming interface (API). This would allow external software developers to integrate these models into their own applications, the report said.
Microsoft and OpenAI have not yet commented on the report.
