Phi-4
model14B parameter dense Transformer with emphasis on synthetic data for complex reasoning. Surpasses its teacher model on STEM benchmarks. Also released as Phi-4 Mini (3.8B, 128K context) and Phi-4 Reasoning (chain-of-thought, April 2025).
Phi-4 Reasoning Plus achieved 78% on AIME 2025, beating QwQ-32B despite being less than half the size. AA Intelligence Index: 10. MIT License.
Model Details
Architecture DENSE
Parameters 14B
Variants
| Name | Parameters | Notes |
|---|---|---|
| Phi-4 | 14B | — |
| Phi-4 Mini | 3.8B | 128K context |
| Phi-4 Reasoning | 14B | Chain-of-thought, April 2025 |
Paper
arXiv: 2412.08905