[whatmodel.sh](../../README.md) / Llama 3.2 3B Instruct
# Llama 3.2 3B Instruct
Meta · released 2024-09 · 3.21B params · llama3.2
**score 40.4** ±0.5 · #557 of 732 · from 65 benchmark results across 17 sources
Score within the [LMArena (text)](https://whatmodel.sh/?set=lmarena-text&score=set) comparison set only.
## What it's good at
● marks Llama 3.2 3B Instruct among every model on that task or test, worst to best.
task vs the others rank counts tests
agent tasks ▁▁▃●█▄▅▅▄▄▂▁ #285 of 313 2.1% 3
coding ▁▁▁▃●▇█▇▅▄▃▁ #380 of 469 10% 7
chat & writing ▁▁▁▃●▄▆█▆▇▇▃ #341 of 402 6.6% 4
math ▁▁▃●█▆▄█▄▂▁▁ #369 of 431 22% 7
agent tasks · #285 of 313 · 2.1% of its score
Beats [Magistral Small 1.0 (2506)](magistral-small-2506.md), [Bielik 11B v2.3 Instruct](bielik-11b-v2-3-instruct.md), [Mixtral 8x22B Instruct v0.1](mixtral-8x22b-instruct-v0-1.md) here, though they rank higher overall. Trails [Granite 4.0 H 350M](granite-4-0-h-350m.md), [Qwen3 0.6B](qwen3-0-6b.md), [Qwen2.5 1.5B Instruct](qwen2-5-1-5b.md) here, though they rank lower overall.
test result vs the others in its run difficulty counts why
BFCL V4 (December 2025 archive) 21.9% ▁▁▁▂●█▅▅▆▂▂▂ beats 8/77 hard · avg 31% 0.8% typical
MASSIVE-Agents AST (10k, zero-sho… 4.3% █····●██·███ beats 1/6 very hard · avg 15% 0.8% sharp · too hard for it
BFCL v3 45.9% ▁▁▁▁▁▃●█▅▄▁▁ beats 25/114 medium · avg 51% 0.6% blunt · at its level
coding · #380 of 469 · 10% of its score
Beats [LFM2.5 8B-A1B](lfm2-5-8b-a1b.md), [Qwen3-VL 2B Thinking](qwen3-vl-2b-thinking.md), [DeepSeek R1 Distill Qwen 7B](deepseek-r1-distill-qwen-7b.md) here, though they rank higher overall. Trails [Qwen2.5 1.5B Instruct](qwen2-5-1-5b.md), [WizardLM 13B V1.2](wizardlm-13b-v1-2.md) here, though they rank lower overall.
test result vs the others in its run difficulty counts why
Aider code editing (133 Exercism … 26.3% ▁·●▂▄▁▆▄█▆▂▂ beats 1/62 medium · avg 51% 3.8% sharp · at its level
HumanEval 45.1% ▄·▄█·●·▄···█ beats 4/7 easy · avg 78% 3.1% sharp · at its level
MBPP 40.2% ▄···▄·▄▄●··█ beats 4/7 medium · avg 61% 1.5% at its level
CodeReviewQA: solution identifica… 7.6% ▃··▁●▂▃▂█▃▄▂ beats 7/47 medium · avg 51% 0.5% sharp · noisy
CodeReviewQA: automated code refi… 25.9% ▁▁▁▁▁▁▂▂●█▅▄ beats 16/47 hard · avg 32% 0.5% typical
CodeReviewQA: change localisation… 0.4% ●▂▁▁▂▂▂▃▅▅▅▇ beats 6/47 hard · avg 33% 0.4% sharp · noisy
CodeReviewQA: change type recogni… 78.8% ▁▁▁······▁▃● beats 44/47 easy · avg 67% 0.0% blunt · at its level · noisy
chat & writing · #341 of 402 · 6.6% of its score
Beats [OpenThinker3 7B](openthinker3-7b.md), [Code Llama 70B Instruct](codellama-70b-instruct.md), [Gemma 4 E2B](gemma-4-e2b.md) here, though they rank higher overall. Trails [Granite 4.0 H 350M](granite-4-0-h-350m.md), [LFM2 350M](lfm2-350m.md), [Gemma 3 1B](gemma-3-1b.md) here, though they rank lower overall.
test result vs the others in its run difficulty counts why
LMArena (text) 1166.8 ▁▁▁▂●▃▃▆▆▇█▃ beats 47/385 avg 1260 2.6% blunt
IFEval mean of four measures 71.4% ●··▄▄·▄····▄ beats 0/5 easy · avg 84% 1.7% at its level
Arabic IFEval prompt-level (Incep… 19.1% ▁·▁▁▂▁▄●▆▄█▄ beats 10/38 medium · avg 46% 1.4% sharp
IFEval (aggregation unspecified) 51.0% ▄····█·▄··●█ beats 5/7 easy · avg 65% 1.0% at its level
math · #369 of 431 · 22% of its score
Beats [GPT-3.5 Turbo (0125)](gpt-3-5-turbo-0125.md), [DBRX Instruct](dbrx-instruct.md), [Command R Plus (04-2024)](command-r-plus-04-2024.md) here, though they rank higher overall. Trails [OPT-1.3B](opt-1-3b.md), [OPT-2.7B](opt-2-7b.md), [Gemma 2B (base)](gemma-2b-base.md) here, though they rank lower overall.
test result vs the others in its run difficulty counts why
GSM8K 75.6% ●·█▄▄█···█·▄ beats 0/10 very easy · avg 90% 9.0% sharp · at its level · several runs
MGSM 44.6% ●▃▃█·▅···▃·▃ beats 1/10 easy · avg 72% 5.3% sharp · at its level
MATH (subset unspecified) 31.4% ▄···▄▄▄▄··●▄ beats 6/7 medium · avg 49% 2.6% at its level
MATH-500 41.2% ●····▃·█···▃ beats 0/5 easy · avg 79% 1.8% at its level · several runs
MATH 46.7% █▄●█···▄▄▄·▄ beats 3/10 medium · avg 61% 1.7% at its level
AIME 2024 6.7% ▁▁▁●▁▃▁▅███▃ beats 6/85 medium · avg 54% 1.5% sharp · at its level · noisy · several …
GSM8K loose substring accuracy 75.2% █·●··███···█ beats 1/5 easy · avg 83% 0.2% blunt
other tests (26)
test result vs the others in its run difficulty counts why
MMLU 61.8% ●▄█▄▄▄·▄▄··▄ beats 1/10 easy · avg 75% 8.9% at its level · several runs
BIG-Bench Hard 55.4% █●▄·█▄·▄▄··▄ beats 2/10 medium · avg 63% 7.5% sharp · several runs
ARC Challenge 76.1% ●▂▂█▂···▄··▂ beats 0/10 very easy · avg 89% 6.7% sharp · at its level · several runs
MMLU-Pro 39.2% █●▄·█▄▄▄·▄·▄ beats 2/10 medium · avg 44% 5.1% several runs
OpenBookQA 72.6% ●··▄█▂·▂···▄ beats 0/10 easy · avg 85% 4.0% at its level · several runs
GPQA (subset unspecified) 24.3% ●··▄▂··▂▂··▄ beats 2/10 hard · avg 37% 3.0% sharp · too hard for it · several runs
GPQA Diamond 30.6% ██····██●··█ beats 4/5 medium · avg 52% 2.9% sharp · several runs
WinoGrande 53.2% ●····▃▃█▃·▅▅ beats 0/10 medium · avg 81% 2.6% at its level · several runs
HellaSwag 77.2% ▃▅█·●▃▃···▃▃ beats 6/10 easy · avg 85% 2.3% at its level · several runs
CPTUB average (SpeakLeash, 0-5 ju… 40.0% ▁▂●▂▃█▃▆▅▄▂▃ beats 8/95 medium · avg 65% 2.1% at its level
Multilingual MMLU (Microsoft) 48.1% ●▄▄██··█···▄ beats 1/10 medium · avg 63% 2.0% at its level
Polish EQ-Bench v2 (SpeakLeash) 46.5% ▁··▁▁▂▂●▅▆█▆ beats 22/70 medium · avg 62% 1.5% at its level
BoolQ 71.4% ▃··●··██▃·▃▃ beats 1/10 easy · avg 88% 1.4% at its level · several runs
BRIDGE clinical text, zero-shot a… 22.9% ▁▁▁▁▁●▂▃▄▆█▃ beats 16/97 hard · avg 31% 1.2% blunt
Multi-IF three-turn average 50.9% ●··█▄··▄···▄ beats 0/5 easy · avg 70% 1.1% at its level
IFBench (aggregation unspecified) 20.8% █··●██··█··█ beats 1/5 hard · avg 31% 1.0% typical
ZebraLogic (ZeroEval, 1,000 puzzl… 7.4% ▁▁●▆██▁▁·▁▁▁ beats 4/57 hard · avg 15% 1.0% sharp · too hard for it
Social IQA 68.3% ●·▃▅▅█·▃···▃ beats 0/10 medium · avg 76% 1.0% blunt · at its level · several runs
MMMLU seven languages (Liquid) 47.9% ●·█····█···▄ beats 0/5 medium · avg 59% 0.9% at its level
MGSM six languages, loose (Liquid) 61.7% ●·▄·█···▄··▄ beats 0/5 easy · avg 81% 0.8% at its level
AraGen-v3 3C3H score (Inception, … 15.0% ▁▁·▁▂▂●█▄▅▄▄ beats 10/66 hard · avg 29% 0.7% typical
Open PT-LLM Leaderboard average (… 64.8% ▃▁▁▁▁▁▁▁▂●▅▁ beats 112/276 medium · avg 72% 0.6% at its level · noisy
EuroEval Life in the UK (English … 77.0% ▁▁▁▁▁▁▁▂●█▄▃ beats 56/146 easy · avg 82% 0.4% at its level · noisy
GSM-Plus 38.7% ●·····███·██ beats 0/5 easy · avg 71% 0.3% blunt · at its level
PIQA 68.2% ●·▅·▃█·▃▅··▃ beats 0/10 medium · avg 82% 0.2% blunt · at its level
TruthfulQA MC2 59.2% ●·█▂▂▄····▂▂ beats 0/10 medium · avg 64% 0.1% blunt · at its level
source results share of the score
microsoft/Phi-4-mini-instruct model card 16 36% ▇▇▇▇
IBM Granite 3.0 Language Models technic… 15 30% ▇▇▇▇
LFM2 Technical Report 11 11% ▇
Aider code editing benchmark runs 1 4% ▇
LMArena leaderboard dataset 1 3% ▇
microsoft/Phi-4-mini-reasoning model ca… 3 2% ▇
Microsoft 3 2% ▇
SpeakLeash CPTUB 1 2% ▇
Inception Arabic/Hindi leaderboards 2 2% ▇
SpeakLeash Polish EQ-Bench leaderboard 1 2% ▇
CodeReviewQA: The Code Review Comprehen… 4 1% ▇
Berkeley Function Calling Leaderboard r… 2 1% ▇
BRIDGE Medical Leaderboard 1 1% ▇
WildEval ZeroEval published results 1 1% ▇
MASSIVE-Agents 1 1% ▇
Open Portuguese LLM Leaderboard 1 1% ▇
EuroEval leaderboards 1 0% ▇
prior: pull toward the middle 0% ▇
weighted average = 40.4
Licences: microsoft/Phi-4-mini-instruct model card ([MIT](../../licenses/card-microsoft-phi-4-mini-instruct.txt)) · IBM Granite 3.0 Language Models technical report ([Apache-2.0](https://github.com/ibm-granite/granite-3.0-language-models/blob/5960b7fb3308588bf5f86348278ec9d6015ab647/LICENSE)) · LFM2 Technical Report ([CC BY 4.0](../../licenses/cards-gap-NOTICE.txt)) · Aider code editing benchmark runs ([Apache-2.0](../../licenses/aider-polyglot-runs-Apache-2.0.txt)) · LMArena leaderboard dataset ([CC BY 4.0](https://creativecommons.org/licenses/by/4.0/)) · microsoft/Phi-4-mini-reasoning model card ([MIT](../../licenses/card-microsoft-phi-4-mini-reasoning.txt)) · Microsoft ([CC BY 4.0](../../licenses/report-2503.01743v2.txt)) · SpeakLeash CPTUB ([MIT](../../licenses/hf-leaderboard-spaces-MIT.txt)) · Inception Arabic/Hindi leaderboards ([Apache-2.0](../../licenses/hf-leaderboard-spaces-Apache-2.0.txt)) · SpeakLeash Polish EQ-Bench leaderboard ([Apache-2.0](../../licenses/hf-leaderboard-spaces-Apache-2.0.txt)) · CodeReviewQA: The Code Review Comprehension Assessment for Large Language Models ([CC-BY-4.0](https://creativecommons.org/licenses/by/4.0/)) · Berkeley Function Calling Leaderboard result archive ([Apache-2.0](../../licenses/bfcl-result-archive-v4-Apache-2.0.txt)) · BRIDGE Medical Leaderboard ([MIT](../../licenses/bridge-medical-leaderboard-MIT.txt)) · WildEval ZeroEval published results ([Apache-2.0](../../licenses/zeroeval-published-results-Apache-2.0.txt)) · MASSIVE-Agents ([CC BY 4.0](https://creativecommons.org/licenses/by/4.0/)) · Open Portuguese LLM Leaderboard ([Apache-2.0](../../licenses/hf-leaderboard-spaces-Apache-2.0.txt)) · EuroEval leaderboards ([MIT](https://github.com/EuroEval/leaderboards/blob/1a77f1d2f8ba1ba05ab73287fb090d76bf41b2e0/LICENSE))
## How to run it
### Through an API
api (Cloudflare) · $0.12/M blended · $0.051 in / $0.335 out per 1M tokens · [developers.cloudflare.com](https://developers.cloudflare.com/workers-ai/platform/pricing/)
[price source](https://github.com/cloudflare/cloudflare-docs/blob/b182511e39518f4b3304edd7bb3e0b0d3017b035/src/content/docs/workers-ai/platform/pricing.mdx): Workers AI published marginal USD per 1M tokens after free allocation. Workers plan fees excluded; billed in Neurons. Published USD rates are rounded. Product: Workers AI. API model: @cf/meta/llama-3.2-3b-instruct. Settings: plan=Workers Paid, free allocation=excluded, billing=published marginal USD rates; service bills Neurons. Currency: USD. · [context source](https://github.com/anomalyco/models.dev/blob/a668908cfe7073faa0363699486fc41a3ff52412/models/meta/llama-3.2-3b.toml)
### 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 Llama 3.2 3B Instruct (Q4_K_M, needs ~2 GB, from [huggingface.co/lmstudio-community](https://huggingface.co/lmstudio-community/Llama-3.2-3B-Instruct-GGUF/tree/c91307b5cf18c8106b1f8a6218c26ae4dbfee472))
```sh
llama cli -hf lmstudio-community/Llama-3.2-3B-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/Llama-3.2-3B-Instruct-GGUF:Q4_K_M
```
Other versions: Q8_0 (~3.4 GB) `llama cli -hf lmstudio-community/Llama-3.2-3B-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 Llama 3.2 3B Instruct (Q4_K_M, needs ~2 GB, from [huggingface.co/lmstudio-community](https://huggingface.co/lmstudio-community/Llama-3.2-3B-Instruct-GGUF/tree/c91307b5cf18c8106b1f8a6218c26ae4dbfee472))
```sh
lms get https://huggingface.co/lmstudio-community/Llama-3.2-3B-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 (~3.4 GB) `lms get https://huggingface.co/lmstudio-community/Llama-3.2-3B-Instruct-GGUF@Q8_0 -y` · 4bit (~1.8 GB, Mac only) `lms get https://huggingface.co/mlx-community-staging/Llama-3.2-3B-Instruct-mlx-4Bit -y` · 8bit (~3.4 GB, Mac only) `lms get https://huggingface.co/mlx-community/Llama-3.2-3B-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 Llama 3.2 3B Instruct (Q4_K_M, needs ~2.43 GB, from [ollama.com](https://ollama.com/library/llama3.2/tags))
```sh
ollama run llama3.2:3b
```
Other versions: Q8_0 (~4.35 GB) `ollama run llama3.2:3b-instruct-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 Llama 3.2 3B Instruct (4bit, needs ~1.8 GB, Mac only, from [huggingface.co/mlx-community-staging](https://huggingface.co/mlx-community-staging/Llama-3.2-3B-Instruct-mlx-4Bit/tree/84775981df316c5d938308743fe099c72d88a0e8))
```sh
mlx_lm.chat --model mlx-community-staging/Llama-3.2-3B-Instruct-mlx-4Bit
```
3. Or start an OpenAI-compatible server at http://localhost:8080/v1
```sh
mlx_lm.server --model mlx-community-staging/Llama-3.2-3B-Instruct-mlx-4Bit
```
Other versions: 8bit (~3.4 GB, Mac only) `mlx_lm.chat --model mlx-community/Llama-3.2-3B-Instruct-8bit`
Model repository and documentation: [huggingface.co/meta-llama](https://huggingface.co/meta-llama/Llama-3.2-3B-Instruct)
## Published speed measurements
Reference results for the hardware, engine and settings below. These are not verified speeds of the downloads above or predictions for your machine.
hardware precision speed engine and settings
Apple M4 Pro (12-core CPU /… GGUF Q4_K_M (Ol… 115.2 tok/s Ollama v0.21.2 (llama.cpp, Metal)…
Apple M5 Max (40-core GPU),… q4_k_m 161.81 tok/s ollama (version not recorded) · l…
Each speed links to its source, which has the full settings.
## 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
ONNX INT4 AWQ (NVIDIA ModelOp… MMLU 5-shot (ModelOpt-Windows acc… 0.608 → 0.577 94.9%
Each kept value links to its source.
[interactive version](https://whatmodel.sh/model/llama-3-2-3b) · [JSON](https://whatmodel.sh/api?format=json&all=1)