# Phi-4 Reasoning ![Organization](https://img.shields.io/badge/Org-Microsoft-0078d4) ![License](https://img.shields.io/badge/License-Apache%202.0-22c55e) ![Released](https://img.shields.io/badge/Released-Apr%202025-555) ![Reasoning](https://img.shields.io/badge/Type-Reasoning-8250df) > **Quick answer:** Phi-4 Reasoning is Microsoft's April 2025 reasoning-optimised model scoring 97.2% GSM8K, 92.7% MATH, 77.8% GPQA Diamond, 74.3% MMLU-Pro, 53.8% LiveCodeBench, and 73.3% Arena Hard. Apache 2.0. ## At a Glance **Where Phi-4 Reasoning leads** - 97.2% GSM8K — near-perfect grade-school math - 92.7% MATH — excellent competition math - 77.8% GPQA Diamond — solid graduate-level science - 53.8% LiveCodeBench — competitive coding - Apache 2.0 — fully open **Where it lags** - 73.3% Arena Hard — moderate instruction quality - 74.3% MMLU-Pro — moderate academic breadth - Superseded by Phi-4 Reasoning Plus on most metrics **Best for:** Open-source math and reasoning tasks; Apache 2.0 reasoning-first pipelines; teams where Phi-4 Reasoning Plus is too large/expensive. ## What Phi-4 Reasoning Is Phi-4 Reasoning is the chain-of-thought reasoning variant of Microsoft's Phi-4 (14B) model, released April 2025. It is the direct predecessor to Phi-4 Reasoning Plus, extending Phi-4's base capabilities with reasoning training (similar to o1-mini style chain-of-thought). The 92.7% MATH and 77.8% GPQA Diamond scores show the reasoning training effect — substantially higher than Phi-4 Mini (64% MATH) and competitive with models several times larger. ## Specifications | Field | Value | |---|---| | Organization | Microsoft | | License | Apache 2.0 | | HuggingFace | microsoft/Phi-4-reasoning | | Release date | April 30, 2025 | | Modality | Text only | ## Pricing Open weights under Apache 2.0 — self-host at no cost. Available via Azure AI Foundry. ## Public Benchmark Scores | Benchmark | Score | Source | Date | |---|---|---|---| | [GSM8K](/benchmarks/gsm8k) | 97.2% | Benchgen evaluation | 2025-04 | | [MATH](/benchmarks/math) | 92.7% | Benchgen evaluation | 2025-04 | | [GPQA Diamond](/benchmarks/gpqa-diamond) | 77.8% | Benchgen evaluation | 2025-04 | | [MMLU-Pro](/benchmarks/mmlu-pro) | 74.3% | Benchgen evaluation | 2025-04 | | [LiveCodeBench](/benchmarks/livecodebench) | 53.8% | Benchgen evaluation | 2025-04 | | [Arena Hard](/benchmarks/arena-hard-v2) | 73.3% | Benchgen evaluation | 2025-04 | ## Phi-4 Reasoning vs Alternatives | Model | MATH | GPQA Diamond | MMLU-Pro | License | |---|---|---|---|---| | **Phi-4 Reasoning** | 92.7% | 77.8% | 74.3% | Apache 2.0 | | [Phi-4 Reasoning Plus](/models/microsoft/phi-4-reasoning-plus) | — | — | 76% | MIT | | [Phi-4](/models/microsoft/phi-4) | — | — | — | MIT | | [QwQ-32B](/models/qwen/qwq-32b) | — | — | — | Apache 2.0 | Phi-4 Reasoning vs Phi-4 Reasoning Plus: Reasoning Plus scores 79% Arena Hard (vs 73.3%), 53.1% LiveCodeBench (vs 53.8%), 76% MMLU-Pro (vs 74.3%). Use Reasoning Plus for maximum Phi-4 reasoning capability; Reasoning for a slightly lighter variant. ## Related Models & Benchmarks - Compare: [Phi-4 Reasoning Plus](/models/microsoft/phi-4-reasoning-plus) · [Phi-4](/models/microsoft/phi-4) · [QwQ-32B](/models/qwen/qwq-32b) - Benchmarks: [MATH](/benchmarks/math) · [GPQA Diamond](/benchmarks/gpqa-diamond) · [LiveCodeBench](/benchmarks/livecodebench) ## Frequently Asked Questions

What is Phi-4 Reasoning?

Microsoft's April 2025 reasoning-optimised 14B model scoring 97.2% GSM8K, 92.7% MATH, 77.8% GPQA Diamond. Apache 2.0.

What is the difference between Phi-4 Reasoning and Phi-4 Reasoning Plus?

Phi-4 Reasoning Plus is a stronger variant with higher Arena Hard (79% vs 73.3%) and MMLU-Pro (76% vs 74.3%). Phi-4 Reasoning is the base reasoning variant. Both are Microsoft-built on the Phi-4 architecture. Reasoning Plus is the recommended choice.

--- *Specs from Microsoft's Phi-4 Reasoning release (April 2025) and Benchgen evaluations. Last updated 2026-07-24.*

Phi 4 Reasoning — Evaluations

Evaluations