Microsoft has introduced the most recent addition to its Phi household of generative AI fashions.
Known as Phi-4, the mannequin is improved in a number of areas over its predecessors, Microsoft claims — particularly math downside fixing. That’s partly the results of improved coaching information high quality.
Phi-4 is obtainable in very restricted entry as of Thursday night time: solely on Microsoft’s not too long ago launched Azure AI Foundry improvement platform, and just for analysis functions underneath a Microsoft analysis license settlement.
That is Microsoft’s newest small language mannequin, coming in at 14 billion parameters in dimension, and it competes with different small fashions equivalent to GPT-4o mini, Gemini 2.0 Flash, and Claude 3.5 Haiku. These AI fashions are oftentimes quicker and cheaper to run, however the efficiency of small language fashions has progressively elevated during the last a number of years.
On this case, Microsoft attributes Phi-4’s bounce in efficiency to the usage of “high-quality artificial datasets,” alongside high-quality datasets of human-generated content material and a few unspecified post-training enhancements.
Many AI labs are wanting extra carefully at improvements they’ll make round artificial information and post-training nowadays. Scale AI CEO Alexandr Wang mentioned in a tweet on Thursday that “now we have reached a pre-training information wall,” confirming a number of reviews on the subject within the final a number of weeks.
Notably, Phi-4 is the primary Phi-series mannequin to launch following the departure of Sébastien Bubeck. Beforehand an AI VP at Microsoft and a key determine within the firm’s Phi mannequin improvement, Bubeck left Microsoft in October to hitch OpenAI.