Microsoft has emerged as one of the original supporters of OpenAI, frequently promoting products like Copilot by emphasizing their access to the latest ChatGPT models. Now, the tech giant appears to be shifting toward enhancing its own AI models within its software suite while also creating alternatives to OpenAI’s reasoning models within the "GPT-o" framework.
According to The Information, staff in Microsoft’s AI division have recently finished training a new series of AI models currently under development, dubbed “MAI.” The internal team is optimistic that these models will perform at levels comparable to leading AI frameworks developed by companies like OpenAI and Anthropic.
Under the guidance of its AI head, Mustafa Suleyman, Microsoft is implementing this plan to reduce dependency on OpenAI and establish its own AI framework for Copilot applications. This move isn’t entirely unexpected.
Developing Its Own AI Stack
In the final week of February, Microsoft released new compact language models called Phi-4-multimodal and Phi-4-mini. These models feature multi-modal capabilities, allowing them to process text, speech, and visual inputs, similar to OpenAI's ChatGPT and Google's Gemini.
These AI models are currently accessible to developers via Microsoft’s Azure AI Foundry and third-party platforms such as HuggingFace and NVIDIA API Catalog. Benchmarks provided by Microsoft indicate that the Phi-4 model outperforms Google’s most recent Gemini 2.0 series on several testing criteria.
The company stated, “It is one of the few accessible models capable of effective speech summarization, achieving performance levels comparable to the GPT-4o model,” and anticipates launching its MAI models commercially through its Azure service.
Competition and Willingness to Collaborate
In addition to developing proprietary AI models for Copilot, Microsoft is also exploring partnerships with third-party technologies like DeepSeek, xAI, and Meta. DeepSeek has recently gained attention for offering top performance benchmarks at notably lower development costs. The company has been embraced by various organizations and has claimed a theoretical cost-to-profit ratio exceeding 500% on a daily basis.
Today, we are advancing our AI ambitions with the release of DeepSeek R1 7B & 14B distilled models for Copilot+ PCs via Azure AI Foundry. This marks a significant step in our journey to establish Windows as the premier platform for AI, seamlessly integrating cloud intelligence. pic.twitter.com/QaUYrlMIt6
— Pavan Davuluri (@pavandavuluri) March 3, 2025
While paving the way with its own AI models to replace OpenAI’s GPT underpinning for Copilot, Microsoft is reportedly working on its own reasoning AI models as well. Such efforts would put Microsoft in competition not only with OpenAI's products like GPT-o1 but also with emerging contenders such as DeepSeek, which offer reasoning functions.
This push for in-house reasoning model development has reportedly been accelerated due to tense interactions between Microsoft and OpenAI teams regarding technology sharing. As outlined by The Information, tensions have arisen over OpenAI’s perceived lack of transparency concerning the operations of its AI models, including GPT-o1.
Reasoning models are considered the next essential step in AI development, providing a deeper understanding of queries, enhanced logical deduction, and improved problem-solving skills. Microsoft claims that its Phi-4 model exhibits notable strengths in linguistic, mathematical, and scientific reasoning.
