Leaderboard / Microsoft (Phi)

Phi-4-mini

Microsoft (Phi) · Phi-4 · released 2025-02-26 · data checked 2026-11-01 Needs verification

Phi-4-mini is a 3.8B-parameter dense model from Microsoft (Phi) (United States), released 2025-02-26 under the MIT license. It scores 15.8 on the Small LM Index, #23 of 27 tiny models, with 9 independently measured benchmarks. It needs about 4 GB of memory at 4-bit with an 8K-token context.

Small LM Index

15.8

#23 of 27 scored tiny models

Likely 14.5–17.1 · from 5 of 6 core benchmarks

Parameters

3.8B

Dense

Price per 1M tokens

—

No hosted API on OpenRouter; self-host

Context window

128K

tokens

Benchmarks

Each bar shows this model's score. The tick marks the best tiny model on that benchmark. Sources: 0 reported by Microsoft (Phi), 9 independently measured. How we label sources.

MMLU-Pro Independent Knowledge46.5

Measured by Artificial Analysis. The lab reported 52.8.

Class best: Qwen3.5 4B 79.1

GPQA Diamond Independent Knowledge33.1

Measured by Artificial Analysis.

Class best: Nanbeige4.2 3B 87.4

Humanity's Last Exam Independent Reasoning4.4

Measured by Artificial Analysis.

Class best: Spark-X2.6 4B 17.9

AIME Independent Math3.0

Measured by Artificial Analysis.

Class best: Spark-X2.6 4B 91.2

LiveCodeBench Independent Coding12.6

Measured by Artificial Analysis.

Class best: LFM3 1.7B 73.5

τ²-Bench Independent Agentic8.2

Measured by Artificial Analysis.

Class best: MiniCPM5 2B 97.1

IFBench Independent Instruction following21.1

Measured by Artificial Analysis.

Class best: Spark-X2.6 4B 75.8

MATH-500 Independent Math69.6

Measured by Artificial Analysis.

Class best: MiniCPM5 2B 94.6

AA-LCR Independent Long context15.3

Measured by Artificial Analysis.

Class best: Spark-X2.6 4B 67.3

Not reported: estimated

What the index model expects from this model's other scores. Not a measurement.

SWE-bench Verified≈ 26

Specifications

Lab: Microsoft (Phi) (US · United States)
Architecture: Dense transformer
Total parameters: 3.8B
Active parameters: All (dense)
Size class: Tiny
Context window: 128K tokens
Input modalities: text
Reasoning: Instruct (no thinking)
License: MIT
Openness: Open weights
Memory to run: ~4 GB at 4-bit, 8K context (more sizes)
Hugging Face: 340.5K downloads (30 days) · 881 likes
Measured speed: 46 tokens/s · 0.35s to first token (Artificial Analysis, median across providers)

Weights on Hugging Face ↗

Compare with

Phi-4-mini vs Nemotron Labs Dory 1.0 4B (4B, index 58.6)
Phi-4-mini vs Qwen3.5 4B (4B, index 59.9)
Phi-4-mini vs Qwen3 4B Thinking 2507 (4B, index 59.7)
Phi-4-mini vs Spark-X2.6 4B (4.1B, index 66.5)

Sources

https://huggingface.co/microsoft/Phi-4-mini-instruct

Something wrong here? Report it.

Run it locally

Uses unsloth/Phi-4-mini-instruct-GGUF (Q4_K_M). Swap the tag for another size: BF16, Q2_K, Q2_K_L, Q3_K_M, Q4_K_M, Q5_K_M, Q6_K, Q8_0.

Ollama

ollama run hf.co/unsloth/Phi-4-mini-instruct-GGUF:Q4_K_M

llama.cpp

llama-server -hf unsloth/Phi-4-mini-instruct-GGUF:Q4_K_M

Then open http://localhost:8080 for a chat window.

MLX (Mac)

pip install -U mlx-lm
mlx_lm.chat --model mlx-community/Phi-4-mini-instruct-4bit

vLLM (GPU server)

vllm serve microsoft/Phi-4-mini-instruct

Full-precision weights, for a server GPU.

Open in LM Studio

Memory needed: weights, context cache and runtime buffers.

Quantization	4K context	32K context	128K context
16-bit F16	8.7 GB	12 GB	25 GB
8-bit Q8_0	5.2 GB	8.9 GB	21 GB
4-bit Q4_K_M	3.4 GB	7.2 GB	20 GB

Check it against your GPU, Mac or phone →