Pulse News

**Microsoft Trains New In-House AI Models; Tests Out DeepSeek, Meta in Copilot**

By Aaron Holmes

Source: The Information

Staff in Microsoft's artificial intelligence unit recently finished training a new family of AI models, internally dubbed MAI, that they think can perform nearly as well as leading models from labs like OpenAI and Anthropic, according to a person involved in the effort. The initiative is part of Microsoft AI CEO Mustafa Suleyman's plan to make Microsoft more self-sufficient in AI after primarily relying on OpenAI's models to power its Copilot chatbot apps. Microsoft is also developing so-called reasoning models that could compete with the likes of OpenAI's GPT model, this person said.

Microsoft is now testing out ways to replace OpenAI's models in Copilot with its own homegrown models, as well as with models from other AI labs including XAI, Meta Platforms, and the Chinese startup DeepSeek, this person said. Microsoft is also considering making its MAI models commercially available on its Azure cloud computing service, which would bring its AI lab into even more direct competition with OpenAI and other AI firms.

Suleyman is simultaneously tasked with carrying on the Microsoft-OpenAI partnership, which is set to run through at least 2030 and which guarantees Microsoft a share of OpenAI's revenue and future profits, as well as the rights to reuse its technology. But Suleyman has at times clashed with OpenAI over how the startup delivers technology to Microsoft, arguing that the startup wasn't holding up its end of the bargain by failing to explain how models like GPT work in detail, according to people familiar with the discussions.