Copilot may soon see an increase in Microsoft AI models with a reduced reliance on ChatGPT.

      Microsoft has been an early supporter of OpenAI and has frequently promoted products like Copilot by highlighting their access to the latest ChatGPT models. Now, it appears that Microsoft aims to advance its own AI models within its popular software suite, while simultaneously developing a competitor to OpenAI's reasoning models in the "GPT-o" category.

      According to The Information, the employees within Microsoft's AI division have recently completed training a "new family of AI model" currently under development, codenamed "MAI." The internal team is optimistic that these in-house models will perform nearly as well as leading AI models from OpenAI and Anthropic.

      With Mustafa Suleyman at the helm of its AI operations, Microsoft is initiating this project to reduce its reliance on OpenAI and build its own AI infrastructure for Copilot applications. These developments are not unexpected.

      Steadily building its own stack

      In late February, Microsoft launched new small language models named Phi-4-multimodal and Phi-4-mini. These models feature multi-modal capabilities, enabling them to process text, speech, and visual inputs, similar to OpenAI’s ChatGPT and Google’s Gemini.

      These two new AI models are already accessible to developers through Microsoft’s Azure AI Foundry and various third-party platforms like HuggingFace and the NVIDIA API Catalog. In benchmarks shared by the company, the Phi-4 model outperforms Google’s latest Gemini 2.0 models across multiple testing metrics.

      “It is among a select group of open models to successfully implement speech summarization and achieve performance levels comparable to the GPT-4o model,” Microsoft noted in its blog post. The company anticipates launching its "MAI" models commercially through its Azure service.

      Rivalry, and openness to rivals

      In addition to testing its own AI models for Copilot, Microsoft is also looking into third-party alternatives such as DeepSeek, xAI, and Meta. DeepSeek recently gained attention for providing a high-performance benchmark at significantly lower development costs and has already been embraced by multiple companies, claiming a theoretical cost-to-profit ratio exceeding 500% daily.

      Today, we are enhancing our AI initiatives with the launch of DeepSeek R1 7B & 14B distilled models for Copilot+ PCs via Azure AI Foundry. This marks the next step in our mission to establish Windows as the leading platform for AI, seamlessly incorporating intelligence from the cloud to…

      Additionally, while Microsoft develops its own AI models to supplant OpenAI’s GPT framework for Copilot, the company is also reportedly creating its own reasoning AI models. This positions Microsoft in competition not only with OpenAI products like GPT-o1 but also with emerging competitors like DeepSeek, both of which provide reasoning capabilities.

      The development of in-house reasoning models has been accelerated due to strained relations between Microsoft and OpenAI teams concerning technology sharing. As reported by The Information, Suleyman and OpenAI have encountered disagreements over the latter's lack of transparency regarding the workings of its AI models such as GPT-o1.

      Reasoning models are considered the next significant advancement in AI development, as they provide a more refined understanding of queries, logical deduction, and improved problem-solving abilities. Microsoft also claims that its Phi-4 model exhibits superior language, mathematical, and visual science reasoning capabilities.