Microsoft has unveiled the latest member of its Phi collection of generative AI models.

Christened Phi-4, the model boasts significant enhancements over its predecessors, particularly in solving mathematical problems, according to Microsoft. This is largely attributed to the improved quality of the training data used.

As of Thursday evening, access to Phi-4 is extremely limited. It is exclusively available on Microsoft’s newly launched Azure AI Foundry development platform and solely for research under a Microsoft research license agreement.

Phi-4 is Microsoft’s most recent compact language model, with an impressive 14 billion parameters, putting it in competition with other small models like GPT-4o mini, Gemini 2.0 Flash, and Claude 3.5 Haiku. These AI models are often more cost-effective and faster to operate, and their performance has seen a steady increase over recent years.

In this instance, Microsoft attributes the performance leap of Phi-4 to the utilization of “high-quality synthetic datasets”, supplemented with high-quality human-generated data sets and some post-training enhancements, the specifics of which remain undisclosed.

Many AI labs are currently focusing on potential innovations around synthetic data and post-training. Scale AI CEO, Alexandr Wang, confirmed in a Thursday tweet that “we have reached a pre-training data wall,” echoing several recent discussions on the subject.

Intriguingly, Phi-4 is the inaugural Phi-series model to be released after the exit of Sébastien Bubeck, a former AI VP at Microsoft and instrumental in the development of the Phi model. Bubeck transitioned from Microsoft to OpenAI in October.