← All models

Phi-4-reasoning

Microsoft · 2025-04-30 · 14.7B parameters

Phi-4-reasoning is Microsoft Research's open-weight 14B reasoning model, obtained by supervised fine-tuning of Phi-4 on more than 1.4 million prompt-response pairs with long reasoning traces generated by o3-mini. Its RoPE base frequency was doubled to extend Phi-4's 16K context to 32,768 tokens, and it emits a <think> reasoning block followed by a summarized solution. The technical report describes this checkpoint as SFT-only; the reinforcement-learning-enhanced variant is the separately released Phi-4-reasoning-plus.

Benchmark scores