[whatmodel.sh](../../README.md) / Phi-4-mini 3.8B
# Phi-4-mini 3.8B
Microsoft · released 2025-02 · 3.84B params · mit
**score 45.2** ±0.6 · #472 of 732 · from 40 benchmark results across 10 sources
Score within the [LMArena (text)](https://whatmodel.sh/?set=lmarena-text&score=set) comparison set only.
## What it's good at
● marks Phi-4-mini 3.8B among every model on that task or test, worst to best.
task vs the others rank counts tests
agent tasks ▁●▃▆█▄▅▅▄▄▂▁ #307 of 313 2.4% 3
coding ▁▁▁▃▅●█▇▅▄▃▁ #317 of 469 7.2% 4
chat & writing █··········● #1 of 2* 0.6% 1
math ▁▁▃▅●▆▄█▄▂▁▁ #306 of 431 22% 5
agent tasks · #307 of 313 · 2.4% of its score
Beats [QwQ-32B-Preview](qwq-32b-preview.md), [DeepSeek-Coder-V2 Lite Instruct (16B-A2.4B)](deepseek-coder-v2-lite-instruct.md) here, though they rank higher overall. Trails [Granite 4.0 350M](granite-4-0-350m.md), [Granite 4.0 H 350M](granite-4-0-h-350m.md), [Qwen2.5 0.5B Instruct](qwen2-5-0-5b.md) here, though they rank lower overall.
test result vs the others in its run difficulty counts why
BFCL v3 28.2% ▁▁●▁▁▃▅█▅▄▁▁ beats 6/114 medium · avg 51% 1.0% blunt · at its level
BFCL V4 (2025-11-03 archive, mode… 21.7% ●··▅▃▃·▅·▃·█ beats 0/10 hard · avg 33% 0.8% at its level
BFCL (version not stated) (maker) 70.3% ▄▆▄●·█▆·▄··▂ beats 7/17 easy · avg 76% 0.6% blunt · at its level · small run
coding · #317 of 469 · 7.2% of its score
Beats [Solar Open 100B (102B-A12B)](solar-open-100b.md), [Qwen3-VL 30B-A3B Instruct](qwen3-vl-30b-a3b-instruct.md), [Qwen3 4B Instruct 2507](qwen3-4b-instruct-2507.md) here, though they rank higher overall. Trails [DeepSeek LLM 67B Chat](deepseek-llm-67b-chat.md), [WizardLM 70B V1.0](wizardlm-70b-v1-0.md) here, though they rank lower overall.
test result vs the others in its run difficulty counts why
HumanEval (maker) 74.4% ▁▁▁▁▂▂▆▆●▅▂▁ beats 155/205 easy · avg 82% 2.6% sharp · at its level · small run
HumanEval+ (maker) 68.3% ▁·▁▁▂▃▄▄●▅▄▃ beats 74/145 easy · avg 78% 2.4% sharp · at its level · small run
MBPP (maker) 65.3% ▁▁▁▄▅█▆●▄▁▂▁ beats 61/86 easy · avg 74% 1.3% at its level · small run
MBPP+ (maker) 63.8% ▁··▁▁▃▄●▆▄▁▁ beats 58/120 easy · avg 70% 0.9% at its level · small run
chat & writing · #1 of 2 · 0.6% of its score
test result vs the others in its run difficulty counts why
IFEval (aggregation unspecified) … 70.1% ▁▂▄▂●▄▆██▅▂▁ beats 32/145 easy · avg 74% 0.6% at its level · small run
math · #306 of 431 · 22% of its score
Beats [Molmo 2 8B](molmo-2-8b.md), [Llama 3.1 Nemotron 70B Instruct](llama-3-1-nemotron-70b.md), [GPT-4 (0613)](gpt-4-0613.md) here, though they rank higher overall. Trails [DeepSeek R1 Distill Qwen 1.5B](deepseek-r1-distill-qwen-1-5b.md), [Gemma 7B (base)](gemma-7b-base.md), [Mistral 7B v0.1 (base)](mistral-7b-v0-1.md) here, though they rank lower overall.
test result vs the others in its run difficulty counts why
MGSM (maker) 63.9% ▅▃▃█·●···▃·▃ beats 7/10 easy · avg 72% 9.1% sharp · at its level
GSM8K (maker) 88.6% █·█▄▄█···●·▄ beats 8/10 very easy · avg 90% 7.7% sharp
MATH (maker) 64.0% █▄██···▄▄●·▄ beats 9/10 medium · avg 61% 3.1% at its level
AIME 2024 (maker) 10.0% ▁▁▁●▁▃▁▅███▃ beats 9/85 medium · avg 44% 1.3% sharp · at its level · noisy
MATH-500 (maker) 71.8% ▁▃▂●▄▇▅▇█▄▂▁ beats 14/85 easy · avg 84% 0.7% noisy
other tests (27)
test result vs the others in its run difficulty counts why
ARC Challenge (maker) 83.7% ▂▂▂●▂···▄··▂ beats 6/10 very easy · avg 89% 7.6% sharp
CLIcK Korean (Daekeun Kim) 42.8% ●▄▄▄█▄▂·▄█·▄ beats 0/23 medium · avg 56% 7.2% at its level
MMLU (maker) 67.3% █▄█▄●▄·▄▄··▄ beats 6/10 easy · avg 75% 6.6% at its level
HAE-RAE 1.0 (Daekeun Kim) 44.5% ●▄▄▂▂█▂▄▂█▂▄ beats 1/22 medium · avg 61% 6.5% at its level
BIG-Bench Hard (maker) 70.4% ██▄·█▄·●▄··▄ beats 8/10 medium · avg 71% 5.4% sharp · at its level
OpenBookQA (maker) 79.2% ▂··●█▂·▂···▄ beats 1/10 easy · avg 85% 4.2% at its level
MMLU-Pro (maker) 52.8% █▄▄·█▄▄●·▄·▄ beats 8/10 medium · avg 54% 3.8% at its level
CPTUB average (SpeakLeash, 0-5 ju… 43.4% ▁▂●▂▃█▃▆▅▄▂▃ beats 11/95 medium · avg 65% 3.5% at its level
Multilingual MMLU (Microsoft) (ma… 49.3% █●▄██··█···▄ beats 2/10 medium · avg 63% 3.3% at its level
Polish EQ-Bench v2 (SpeakLeash) 50.8% ▁··▁▁▂▂▆●▆█▆ beats 25/70 medium · avg 62% 2.7% at its level
GPQA Diamond (maker) 36.9% ▁▁▁▁▂●▅▅███▄ beats 38/322 hard · avg 49% 2.6% sharp
BRIDGE clinical text, zero-shot a… 24.5% ▁▁▁▁▁▂●▃▄▆█▃ beats 17/97 hard · avg 31% 2.2% blunt · at its level
HellaSwag (maker) 69.1% ●▅█·▃▃▃···▃▃ beats 0/10 easy · avg 82% 1.6% at its level
KCL-Essay vanilla 6.0% ●···▁▃▂█▂▃▃▂ beats 0/23 hard · avg 17% 1.5% typical
WinoGrande (maker) 67.0% ▃····▃▃█●·▅▅ beats 6/10 medium · avg 71% 1.4% at its level
Social IQA (maker) 72.5% ▃·▃●▅█·▃···▃ beats 2/10 medium · avg 76% 1.3% blunt · at its level
BoolQ (maker) 81.2% ▃··▃··█●▃·▃▃ beats 7/10 easy · avg 83% 1.1% at its level
Open PT-LLM Leaderboard average (… 66.5% ▃▁▁▁▁▁▁▁▂●▅▁ beats 127/276 medium · avg 72% 1.0% at its level · noisy
GPQA (subset unspecified) (maker) 25.2% ●··▄▂··▂▂··▄ beats 3/10 very hard · avg 32% 0.9% sharp · too hard for it · noisy
BigCodeBench (Complete) (maker) 43.0% ▂▁▁▃▂▅▇●▅▂▂▂ beats 35/60 medium · avg 49% 0.8% at its level · small run
KMMLU-HARD (Daekeun Kim, zero-sho… 25.3% ●···▁▁▁▂▃·▆▂ beats 5/23 very hard · avg 29% 0.6% sharp · too hard for it
KCL-MCQA vanilla 22.0% ●···█▂▆▆██▆▂ beats 0/23 very hard · avg 29% 0.6% too hard for it
BigCodeBench (Instruct) (maker) 33.8% ▁▁··▁▁▂▅●▅▅▅ beats 20/49 hard · avg 34% 0.5% at its level · small run
PIQA (maker) 77.6% ▃·▅·▃●·▃▅··▃ beats 4/10 medium · avg 82% 0.4% blunt · at its level
Vectara hallucination leaderboard… 96.6% ▁▁▂▃▄▄█●▃▂▂▁ beats 96/140 very easy · avg 94% 0.3% blunt · maxed out · noisy
TruthfulQA MC2 (maker) 66.4% ▂·█▂●▄····▂▂ beats 6/10 medium · avg 64% 0.2% blunt · at its level
Vectara hallucination leaderboard… 76.5% ●▂▁▂▅█▂▃▄▃▁▁ beats 1/98 very easy · avg 90% 0.0% blunt
source results share of the score
microsoft/Phi-4-mini-instruct model car… 16 58% ▇▇▇▇▇▇▇
Daekeun Kim Korean LLM benchmark runs 3 14% ▇▇
Microsoft (submitted / maker-reported) 11 14% ▇▇
SpeakLeash CPTUB 1 3% ▇
SpeakLeash Polish EQ-Bench leaderboard 1 3% ▇
BRIDGE Medical Leaderboard 1 2% ▇
Korean Canonical Legal Benchmark: Towar… 2 2% ▇
Berkeley Function Calling Leaderboard r… 2 2% ▇
Open Portuguese LLM Leaderboard 1 1% ▇
Vectara hallucination leaderboard 2 0% ▇
prior: pull toward the middle 0% ▇
weighted average = 45.2
Licences: microsoft/Phi-4-mini-instruct model card ([MIT](../../licenses/card-microsoft-phi-4-mini-instruct.txt)) · Daekeun Kim Korean LLM benchmark runs ([MIT](../../licenses/korean-personal-benchmark-runs-MIT.txt)) · Microsoft ([CC BY 4.0](../../licenses/report-2503.01743v2.txt)) · SpeakLeash CPTUB ([MIT](../../licenses/hf-leaderboard-spaces-MIT.txt)) · SpeakLeash Polish EQ-Bench leaderboard ([Apache-2.0](../../licenses/hf-leaderboard-spaces-Apache-2.0.txt)) · BRIDGE Medical Leaderboard ([MIT](../../licenses/bridge-medical-leaderboard-MIT.txt)) · Korean Canonical Legal Benchmark: Toward Knowledge-Independent Evaluation of LLMs’ Legal Reasoning Capabilities ([CC BY 4.0](../../licenses/cards-final-NOTICE.txt)) · Berkeley Function Calling Leaderboard result archive ([Apache-2.0](../../licenses/bfcl-result-archive-v4-Apache-2.0.txt)) · Open Portuguese LLM Leaderboard ([Apache-2.0](../../licenses/hf-leaderboard-spaces-Apache-2.0.txt)) · Vectara hallucination leaderboard ([Apache-2.0](https://www.apache.org/licenses/LICENSE-2.0))
## How to run it
### Through an API
api (Microsoft Azure) · $0.13/M blended · $0.075 in / $0.3 out per 1M tokens · [azure.microsoft.com](https://azure.microsoft.com/en-us/pricing/details/ai-foundry-models/microsoft)
[price source](https://azure.microsoft.com/en-us/pricing/details/ai-foundry-models/microsoft): USD per 1M tokens, Azure AI Foundry Models pay-as-you-go (Microsoft models table), 'Phi-4-mini' (Azure model id Phi-4-mini-instruct, GA), 128K context: input $0.075, output $0.30, region US East. Fine-tuned hosting/training rates excluded. Currency: USD. Microsoft Azure official pricing page (price cells carry the amounts in the page's own data), observed 2026-09-25. · [context source](https://github.com/anomalyco/models.dev/blob/a668908cfe7073faa0363699486fc41a3ff52412/models/microsoft/phi-4-mini.toml): model catalogue context ceiling (not a promise by a hosted API)
### llama.cpp (recommended)
1. Install llama.cpp. Homebrew and winget installs name the commands llama-cli and llama-server instead. [Other ways to install](https://github.com/ggml-org/llama.cpp/blob/master/docs/install.md)
```sh
# macOS / Linux
curl -LsSf https://llama.app/install.sh | sh
# Windows
irm https://llama.app/install.ps1 | iex
```
2. Run Phi-4-mini 3.8B (Q4_K_M, needs ~2.5 GB, from [huggingface.co/lmstudio-community](https://huggingface.co/lmstudio-community/Phi-4-mini-instruct-GGUF/tree/009f2b81869a0afd6b9aae23d29fa7b7ee2dcac9))
```sh
llama cli -hf lmstudio-community/Phi-4-mini-instruct-GGUF:Q4_K_M
```
3. Or start a local server and open the address it prints (http://localhost:9931)
```sh
llama serve -hf lmstudio-community/Phi-4-mini-instruct-GGUF:Q4_K_M
```
Other versions: Q8_0 (~4.1 GB) `llama cli -hf lmstudio-community/Phi-4-mini-instruct-GGUF:Q8_0`
### LM Studio
1. Install LM Studio. Open the app once before using the lms command. [Download it (macOS / Windows / Linux)](https://lmstudio.ai/download)
2. Run Phi-4-mini 3.8B (Q4_K_M, needs ~2.5 GB, from [huggingface.co/lmstudio-community](https://huggingface.co/lmstudio-community/Phi-4-mini-instruct-GGUF/tree/009f2b81869a0afd6b9aae23d29fa7b7ee2dcac9))
```sh
lms get https://huggingface.co/lmstudio-community/Phi-4-mini-instruct-GGUF@Q4_K_M -y
```
3. Then run lms ls to see its name, and lms chat followed by that name
Other versions: Q8_0 (~4.1 GB) `lms get https://huggingface.co/lmstudio-community/Phi-4-mini-instruct-GGUF@Q8_0 -y` · 4bit (~2.2 GB, Mac only) `lms get https://huggingface.co/mlx-community/Phi-4-mini-instruct-4bit -y` · 8bit (~4.1 GB, Mac only) `lms get https://huggingface.co/mlx-community/Phi-4-mini-instruct-8bit -y`
### Ollama
1. Install Ollama. [Download it (macOS / Windows)](https://ollama.com/download)
```sh
# Linux
curl -fsSL https://ollama.com/install.sh | sh
```
2. Run Phi-4-mini 3.8B (Q4_K_M, needs ~2.8 GB, from [ollama.com](https://ollama.com/library/phi4-mini/tags))
```sh
ollama run phi4-mini:3.8b
```
Other versions: Q8_0 (~5.11 GB) `ollama run phi4-mini:3.8b-q8_0`
### MLX (Mac)
1. Install MLX (Mac). Apple silicon Macs only, text only.
```sh
# with uv
uv tool install mlx-lm
# or
pip install mlx-lm
```
2. Run Phi-4-mini 3.8B (4bit, needs ~2.2 GB, Mac only, from [huggingface.co/mlx-community](https://huggingface.co/mlx-community/Phi-4-mini-instruct-4bit/tree/ac1c269cb4222a4e136a3d09edad301056c1f36a))
```sh
mlx_lm.chat --model mlx-community/Phi-4-mini-instruct-4bit
```
3. Or start an OpenAI-compatible server at http://localhost:8080/v1
```sh
mlx_lm.server --model mlx-community/Phi-4-mini-instruct-4bit
```
Other versions: 8bit (~4.1 GB, Mac only) `mlx_lm.chat --model mlx-community/Phi-4-mini-instruct-8bit`
Model repository and documentation: [huggingface.co/microsoft](https://huggingface.co/microsoft/Phi-4-mini-instruct)
## Quantization evidence
How much of the full-precision result a quantized version kept, measured within one study. Rows from the same study share a table and aren't independent. They don't describe every download of that precision.
precision test full → quantized kept
FP8 W8A8 dynamic (weights sym… MMLU with chain-of-thought prompt… 0.67 → 0.67 99.9%
FP8 W8A8 dynamic (weights sym… MMLU-Pro (lm-evaluation-harness, … 0.468 → 0.456 97.5%
Each kept value links to its source.
[interactive version](https://whatmodel.sh/model/phi-4-mini) · [JSON](https://whatmodel.sh/api?format=json&all=1)