Microsoft is one of the first donors in Openai, and has repeatedly stocked products like Copilot by praising their access to the latest Chatgpt models. Now, it seems that Microsoft seeks to push his own AI models in the suite of popular software, while developing a rival of Openai reasoning models in the “GPT-O” family.
About InformationMicrosoft AI unit employees recently concluded the training of “a new IA model family” which is currently developing under the code name “May”. Internally, the team hopes that these internal models work almost as well as the best IA models from Openai and Anthropic.
Under the supervision of his AI chief, Mustafa Suleyman, Microsoft launches this initiative to reduce his dependence on Openai and develop his own AI stack for Co -pilot applications. Developments are not surprising.
Regularly build your own battery
During the last week of February, Microsoft introduced new models of small languages called Phi-4-Multimodal and Phi-4-Mini. They are delivered with multimodal capacities, which means that they can treat text, speech and vision as entry formats, just like the Openai Chatppt and Google Gemini.
These two new AI models are already available for developers via the Azure AI foundry of Microsoft and third-party platforms such as Huggingface and the API Nvidia catalog. In the landmarks shared by the company, the Phi-4 model is already ahead of The latest models in the Google Gemini 2.0 series On several test parameters.
“It is one of a few open models to successfully implement the speech summary and reach performance levels comparable to the GPT-4O model,” noted Microsoft in his blog. The company hopes to publish its commercially “May” models via its Azure service.
Rivalry and opening to rivals
In addition to testing internal AI models for Copilot, Microsoft also explores third -party options such as Deepseek, Xai and Meta. Deepseek recently made waves By offering a high performance reference at a considerably lower cost of development. He has already been adopted by many companies and recently claimed A theoretical cost / profit ratio of more than 500% on a daily basis.
Today, we are progressing our AI ambitions with the release of DEEPSEEK R1 7B and 14B distilled models for COPILOT + PCS via Azure AI FOUNDRY. This is the next step in our trip to continue to make Windows the platform for AI, perfectly incorporating the intelligence from the cloud to … pic.twitter.com/qauyrlmit6
– Pavan Davuluri (@Pavandavuluri) March 3, 2025
In addition to developing its own AI models to replace OPENAI’s GPT infrastructure for Copilot, Microsoft will also work on its own reasoning AI models. It would place Microsoft against Openai products such as GPT-O1 as well as Chinese parameters such as Deepseek, which both offer reasoning capacities.
Apparently, work on an internal reasoning model has been accelerated due to tense relationships between Microsoft and OpenAi teams for technology sharing. According to information, Suleyman and Openai disagreed on the lack of transparency of the latter concerning the complex operation of its AI models such as GPT-O1.
The models of reasoning are considered the next border for the development of AI, as they offer a more nuanced understanding of requests, a logical deduction and better capacity for problem solving. Microsoft also says that its Phi-4 model offers linguistic, mathematics and visual reasoning chops.