Model

Phi-4-mini (3.8B)

OPEN

Microsoft · Phi-4 family · released Feb 26, 2025

Compact text-only Phi-4 (128K context) with built-in function calling and 22-language support; MIT-licensed, tops the 3-4B class.

ReasoningCodingVisionFunction callingTool useAgentic

472.4

Elo · rank #322

Parameters

3.8B

Active params

3.8B (dense)

Context

128K tokens

Architecture

Dense decoder-only Transformer with grouped-query attention

License

MIT

Languages

22+

API price (in/out)

$0.075 / $0.3

Modalities

text

Benchmark results

Bar shows position within the tracked field; marker = field best

ARC-ChallengeReasoning83.7%#34

best: Llama 3.1 405B · 96.9%

Arena-HardHuman preference32.8%#50

best: Qwen3-235B-A22B (Non-Thinking) · 96.1%

BIG-Bench HardReasoning70.4%#54

best: ERNIE 4.5 300B-A47B · 94.3%

BFCLAgents70.3%#22

best: SmolLM3 3B (No Thinking) · 92.3%

GPQA DiamondReasoning30.4%#271

best: GPT-6 Astra · 96.0%

GSM8KMath88.6%#49

best: Llama 3.1 405B · 96.8%

HellaSwagReasoning69.1%#128

best: Claude 3 Opus · 95.4%

Humanity's Last ExamReasoning4.2%#139

best: Claude Opus 5.5 · 67.7%

HumanEval+Coding68.3%#20

best: Mistral Small 3.2 (24B) · 92.9%

HumanEvalCoding74.4%#66

best: Claude Opus 4.5 · 99.4%

IFEvalReasoning70.1%#113

best: Gemma 4 26B A4B · 98.5%

LiveCodeBenchCoding19.9%#167

best: DeepSeek-V4-Pro (Think Max) · 93.5%

MATH-500Math64.0%#106

best: GPT-5 · 99.4%

MBPP+Coding63.8%#23

best: Llama 3.1 405B · 88.6%

MBPPCoding65.3%#49

best: Llama-3.3-Nemotron-Super-49B v1 (Reasoning On) · 91.3%

MGSMMath63.9%#38

best: OpenAI o4-mini · 93.7%

MMLU-ProKnowledge52.8%#128

best: Claude Fable 5 · 91.5%

MMLUKnowledge67.3%#165

best: OpenAI o3 · 92.9%

MMMLUKnowledge49.3%#53

best: Gemini 3.1 Pro · 92.6%

PIQAReasoning77.6%#55

best: GPT-4o mini · 93.1%

Social IQaReasoning72.5%#9

best: Apple DCLM-Baseline 7B · 82.9%

TruthfulQAKnowledge66.4%#11

best: Phi-3.5-MoE (16x3.8B, 6.6B active) · 77.5%

WinoGrandeReasoning67.0%#112

best: PaLM 2 · 90.9%

Run it locally

VRAM @ Q4

3 GB

VRAM @ FP16

10 GB

Fits on (Q4)

RTX 3060 12GBRTX 4070 Ti 16GBRTX 3090 24GBRTX 4090 24GBRTX 5090 32GBM4 Pro 48GBM3 Max 128GBM3 Ultra 512GBA100 80GBH100 80GBH200 141GBB200 192GB

Throughput data unavailable.

Quantizations

GGUF · MLX · ONNX

Fine-tune it

Permissive

QLoRA4.4 GB1× RTX 3060 12GB

LoRA9.9 GB1× RTX 3060 12GB

Full fine-tune62.8 GB1× A100 80GB

QLoRA SFT on ~10k samples ≈ $2.22 (1× RTX 3060 12GB)

Plan a fine-tune →

Phi-4 family

Elo progression across releases

Phi-4 (14B)Dec 2024945.3 Phi-4-mini (3.8B)Feb 2025472.4 Phi-4-mini-instructFeb 2025466.4 Phi-4-multimodal (5.6B)Feb 2025743.3 Phi-4-multimodal-instructFeb 2025743.3 Phi-4-reasoningApr 20251230.0 Phi-4-reasoning (14B)Apr 20251230.0

Resources

HFmicrosoft/Phi-4-mini-instruct DOCSazure.microsoft.com/en-us/blog/empowering-innovation-the-next-generation-of-the-phi-family/APIAvailable via Microsoft API

API price $0.075/$0.3 · each benchmark row carries its own source badge (see methodology)