Physical Address
304 North Cardinal St.
Dorchester Center, MA 02124
Microsoft has announced new additions to its Phi family of artificial intelligence models.
Called Phi-4, the model is better in several areas than its predecessor, Microsoft says – especially solving mathematical problems. This is probably the result of better control of study data.
Phi-4 is readily available from Thursday night: only on the recently launched Microsoft Azure AI Foundry development platform, and for research purposes under the Microsoft license agreement.
This is the latest version of Microsoft’s language, coming in at 14 billion in size, and it competes with other smaller versions such as GPT-4o mini, Gemini 2.0 Flash, and Claude 3.5 Haiku. These types of AI are usually faster and cheaper to run, but the performance of micro-languages has increased steadily over the past few years.
In this case, Microsoft says that the Phi-4 jump works thanks to the use of “high-quality data sets,” along with high-quality datasets of human-made products and unspecified changes.
Many AI labs are actively looking for new ways to build around data generated by post training these days. Scale AI CEO Alexandr Wang said in a tweet Thursday that “we’ve reached a wall of previously trained data,” to confirm several reports on this topic in the last few weeks.
In particular, the Phi-4 is the first model of the Phi-series to be launched after the departure of Sébastien Bubeck. Bubeck, former VP of AI at Microsoft and instrumental in the development of the company’s Phi model, on the left Microsoft in October to cooperate with OpenAI.
