Phi 4 reasoning
microsoft/Phi-4-reasoning
Open Source · chat · open-weights
Open
Alert me on changes
Intelligence Index via Artificial Analysis · 0–100, higher is better
License: mit · microsoft/Phi-4-reasoning
AI summary
● machine-written
Microsoft releases Phi-4-reasoning, 14B parameter model for step-by-step logic tasks
Phi-4-reasoning is a 14-billion parameter transformer fine-tuned from Phi-4 to enhance complex reasoning capabilities using supervised fine-tuning and reinforcement learning. It targets math, science, and code reasoning tasks with a 32k context window and is optimized for structured two-part responses (reasoning trace followed by solution). The model achieves strong results on specialized benchmarks such as AIME, OmniMath, and LiveCodeBench.
What's new
- Fine-tuned from Phi-4 using chain-of-thought traces and reinforcement learning
- 32k token context window for extended reasoning tasks
- Optimized for structured reasoning format with two-part output
- Released under MIT license for open-weights access
Best for
Math and science reasoning tasksCode-related reasoning and problem-solvingStep-by-step logic in latency-constrained environmentsStructured reasoning requiring chain-of-thought explanations