Microsoft has launched the newest addition to its Phi family of generative AI fashions.
Known as Phi-4, the model is improved in quite a few areas over its predecessors, Microsoft claims — particularly math draw back fixing. That’s partly the outcomes of improved teaching info prime quality.
Phi-4 is obtainable in very restricted entry as of Thursday evening time: solely on Microsoft’s these days launched Azure AI Foundry enchancment platform, and only for evaluation capabilities under a Microsoft evaluation license settlement.
That’s Microsoft’s latest small language model, coming in at 14 billion parameters in measurement, and it competes with totally different small fashions resembling GPT-4o mini, Gemini 2.0 Flash, and Claude 3.5 Haiku. These AI fashions are oftentimes faster and cheaper to run, nevertheless the effectivity of small language fashions has steadily elevated over the past quite a few years.
On this case, Microsoft attributes Phi-4’s soar in effectivity to utilizing “high-quality synthetic datasets,” alongside high-quality datasets of human-generated content material materials and some unspecified post-training enhancements.
Many AI labs are attempting additional intently at enhancements they’re going to make spherical synthetic info and put up teaching these days. Scale AI CEO Alexandr Wang said in a tweet on Thursday that “we have got reached a pre-training info wall,” confirming quite a few experiences on the topic throughout the remaining quite a few weeks.
Notably, Phi-4 is the first Phi-series model to launch following the departure of Sébastien Bubeck. Beforehand an AI VP at Microsoft and a key decide throughout the agency’s Phi model enchancment, Bubeck left Microsoft in October to affix OpenAI.