Microsoft debuts Phi-4, a new model of artificial intelligence, in a research preview

Microsoft has announced a new addition to its Phi family of productive AI models.

Called Phi-4, the model is improved in fewer areas than before, Microsoft claims – especially for solving mathematical problems. That is partly a result of improved training data quality.

Phi-4 is available in limited access as of Thursday night: only on Microsoft’s newly launched Azure AI Foundry development platform, and only for research purposes under Microsoft’s research license agreement.

This is Microsoft’s latest mini-language model, coming in at 14 billion parameters in size, and competing with other mini-models such as GPT-4o mini, Gemini 2.0 Flash, and Claude 3.5 Haiku. These AI models are often fast and cheap to implement, but the performance of small language models has grown steadily over the past few years.

In this case, Microsoft attributes Phi-4’s performance to the use of “high-quality synthetic datasets,” as well as high-quality datasets of human-generated content and other unspecified post-training enhancements.

Many AI labs are looking more closely at the innovations they can make around synthetic data and post training these days. Scale AI CEO Alexandr Wang said in a tweet Thursday that “we’ve hit a wall of pre-training data,” confirming several reports on the topic in the past few weeks.

Notably, the Phi-4 is the first Phi-series model launched after the departure of Sébastien Bubeck. Previously an AI VP at Microsoft and a key figure in the development of the company’s Phi model, Bubeck left Microsoft in October to join OpenAI.