The latest addition is the Phi-4 Reasoning — a 14 billion-parameter model built by applying a supervised fine-tuning (SFT) algorithm to the Phi-4 base model. The researchers also derived the Phi-4 ...