Microsoft was one of the first backers of OpenAI and has promoted products like Copilot with their access to the latest ChatGPT model. Microsoft appears to be promoting its own AI models within the popular software suite while also developing a competitor to OpenAI reasoning models under the “GPT o” family.
As per The Informationreports that Microsoft’s AI unit has recently completed the training of a “new family of AI models” currently under development under the codename “MAI”. The team hopes that the in-house models will perform as well as top AI models such as OpenAI and Anthropic. Microsoft’s AI chief, Mustafa Suleyman is leading this initiative. The goal is to reduce its dependency on OpenAI, and develop Microsoft’s own AI stack for Copilot. These developments are not unexpected.
Building its own stack
Microsoft introduced new small languages models called Phi-4 Multimodal and Phi-4 Mini in the last week of Feb. They have multi-modal capabilities which means that they can process text as well as speech and vision for input formats. This is similar to OpenAI’s ChatGPT or Google’s Gemini.
These new AI models are available to developers through Microsoft’s Azure AI Foundry as well as third-party platforms like HuggingFace or the NVIDIA API catalog. In benchmarks released by the company, Phi-4 is already ahead of Google’s latest Gemini 2.0 models on multiple test parameter.
Microsoft stated that the Phi-4 model is one of a few open-source models that can implement speech summarization with performance levels comparable to GPT-4o. blog post. The company hopes to release its “MAI’ models commercially through its Azure service.
Rivalry and openness towards rivals
In addition to testing in-house AI for Copilot, Microsoft also explores third-party options, such as DeepSeek xAI and Meta. DeepSeek made waves recently by offering a high-performance benchmark at a significantly lower development cost. It has been adopted by many companies. claimed a theoretical cost to profit ratio of more than 500% on a day-to-day basis.
We are proud to announce the release of DeepSeek 7B and 14B distilled models via Azure AI Foundry for Copilot+ PCs. This is a next step in our journey to make Windows the AI platform, seamlessly integrating cloud intelligence to… pic.twitter.com/QaUYrlMIt6
– Pavan Davuluri (@pavandavuluri)””https://twitter.com/pavandavuluri/status/1896664579465511109?ref_src=twsrc%5Etfw””> Microsoft is reportedly also working on its reasoning AI models. Microsoft has reportedly been developing its own AI models for Copilot to replace OpenAI GPT infrastructure. This would pit Microsoft’s reasoning AI models against OpenAI’s GPT-o1 and Chinese upstarts like DeepSeek.
It appears that the work on a reasoning model in-house has been accelerated due to the strained relationship between Microsoft and OpenAI over technology sharing. According to The Information, Suleyman has been at odds with OpenAI over the latter’s lack transparency about the intricate workings its AI models, such as GPT-1.
The next frontier in AI development is reasoning models, which offer a better understanding of queries, logical deduction, and improved problem solving abilities. Microsoft claims that its Phi-4 reasoning model is more powerful in terms of language, mathematics, and visual sciences.
