Phi-4-mini-reasoning

Phi-4-mini-reasoning is a 3.8-billion-parameter small language model released by Microsoft on April 30, 2025, designed and tested specifically for mathematical reasoning rather than general-purpose chat. It carries a 128k-token context window.

Its training recipe combined large-scale mid-training on distilled long reasoning traces, supervised fine-tuning, rollout DPO, and reinforcement learning with verifiable rewards. The mathematical synthetic corpus was generated using DeepSeek-R1, per the model card. Microsoft reported it outperforming OpenAI's o1-mini and DeepSeek-R1-Distill-Llama-70B on PhD-level math and science benchmarks. It is released under the MIT License.