Microsoft has introduced the latest member of its Phi series of generative AI models, called Phi-4, which brings notable improvements over its predecessors, particularly in solving mathematical problems.
These advancements are attributed to the use of higher-quality training data.
Phi-4 is currently available on a restricted basis, accessible only through Microsoft’s Azure AI Foundry development platform for research purposes under a specialized license agreement.
This new model is a compact language model with 14 billion parameters, positioning it as a competitor to other small models like GPT-4o Mini, Gemini 2.0 Flash, and Claude 3.5 Haiku. Compact models like these are faster and more cost-effective to operate, and their performance has steadily improved in recent years.
Microsoft credits Phi-4's enhanced capabilities to the integration of "high-quality synthetic datasets" with human-generated data, along with undisclosed post-training refinements. The use of synthetic data and post-training optimization has become a growing focus in the AI industry. Scale AI’s CEO, Alexandr Wang, recently noted the challenges posed by limited pre-training data, echoing similar concerns in the field.
Phi-4 also marks a milestone for the Phi series, as it is the first model released after the departure of Sébastien Bubeck. A former Vice President of AI at Microsoft and a pivotal figure in the Phi model development, Bubeck left the company in October to join OpenAI.