# Phi-4 Mini ![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-Feb%202025-555) > **Quick answer:** Phi-4 Mini is Microsoft's February 2025 compact model, scoring 88.6% GSM8K, 83.7% ARC-C, 64.0% MATH, and 52.8% MMLU-Pro. Apache 2.0 — Microsoft's compact model for edge and local deployment. ## At a Glance **Where Phi-4 Mini leads** - 88.6% GSM8K — excellent math for a compact model - 83.7% ARC-C — strong commonsense reasoning - 64.0% MATH — solid competition math - Apache 2.0 — fully open - Optimised for low-resource deployment (mobile, edge) **Where it lags** - 32.8% Arena Hard — lower instruction following quality - 52.8% MMLU-Pro — limited academic breadth - 69.1% HellaSwag — moderate language understanding - Small model: limited capacity vs Phi-4 (14B) **Best for:** Edge deployment and local inference; mobile applications requiring math/reasoning; on-device code assistance. ## What Phi-4 Mini Is Phi-4 Mini is Microsoft's February 2025 compact model — the small sibling of Phi-4 (14B). The Phi series is known for "small data, big training" philosophy: Microsoft invests heavily in curated, high-quality training data to achieve strong reasoning in small models. Phi-4 Mini follows this pattern: 88.6% GSM8K at compact size is strong for the era, reflecting focused math training. The 64.0% MATH score is particularly high for a mini-class model. Apache 2.0 makes it freely deployable for commercial use. ## Specifications | Field | Value | |---|---| | Organization | Microsoft | | License | Apache 2.0 | | HuggingFace | microsoft/Phi-4-mini-instruct | | Release date | February 4, 2025 | | Modality | Text only | ## Pricing Open weights under Apache 2.0 — self-host at no cost. Available via Azure AI Foundry and Ollama/local inference. ## Public Benchmark Scores | Benchmark | Score | Source | Date | |---|---|---|---| | [GSM8K](/benchmarks/gsm8k) | 88.6% | Benchgen evaluation | 2025-02 | | [ARC-C](/benchmarks/ai2-reasoning-challenge) | 83.7% | Benchgen evaluation | 2025-02 | | [MATH](/benchmarks/math) | 64.0% | Benchgen evaluation | 2025-02 | | [MMLU-Pro](/benchmarks/mmlu-pro) | 52.8% | Benchgen evaluation | 2025-02 | | [HellaSwag](/benchmarks/hellaswag) | 69.1% | Benchgen evaluation | 2025-02 | | [Arena Hard](/benchmarks/arena-hard-v2) | 32.8% | Benchgen evaluation | 2025-02 | ## Phi-4 Mini vs Alternatives | Model | GSM8K | MATH | License | Size | |---|---|---|---|---| | **Phi-4 Mini** | 88.6% | 64.0% | Apache 2.0 | Mini | | [Phi-4](/models/microsoft/phi-4) | — | — | MIT | 14B | | [Phi-4 Reasoning Plus](/models/microsoft/phi-4-reasoning-plus) | — | — | MIT | 14B | | [Llama 3.2 3B Instruct](/models/meta/llama-3-2-3b-instruct) | 77.7% | — | Llama 3.2 | 3B | Phi-4 Mini vs Llama 3.2 3B: higher GSM8K (88.6% vs 77.7%) and MATH (64%). For maximum compact math/reasoning under Apache 2.0: Phi-4 Mini. For Llama ecosystem compatibility: Llama 3.2 3B. ## Related Models & Benchmarks - Compare: [Phi-4](/models/microsoft/phi-4) · [Phi-4 Reasoning Plus](/models/microsoft/phi-4-reasoning-plus) · [Llama 3.2 3B Instruct](/models/meta/llama-3-2-3b-instruct) - Benchmarks: [GSM8K](/benchmarks/gsm8k) · [MATH](/benchmarks/math) · [MMLU-Pro](/benchmarks/mmlu-pro) ## Frequently Asked Questions

What is Phi-4 Mini?

Microsoft's February 2025 compact model scoring 88.6% GSM8K, 83.7% ARC-C, 64.0% MATH, and 52.8% MMLU-Pro. Apache 2.0 for edge and local deployment.

Should I use Phi-4 Mini or Phi-4?

Phi-4 (14B) for maximum capability within the Phi family. Phi-4 Mini for edge devices, mobile, or deployments where model size is the primary constraint.

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

Phi 4 Mini — Evaluations

Evaluations