whatmodel.sh / Llama 3.2 3B Instruct

Llama 3.2 3B Instruct

+ contribute
Update model informationAdd a place to run itAdd a test resultPropose a benchmarkCorrect a benchmark scoreEdit data on GitHub

Llama 3.2 3B Instruct is an open-weights language model from Meta, released September 2024. Its published size is 3.21B parameters and a 131k-token context window. Licence: llama3.2.

38.4±0.5#556 of 719 LLMsfrom 65 benchmark results across 17 sources

$ how it compares

The LLMs ranked either side of Llama 3.2 3B Instruct, out of 828 with a score, then more from Meta.

#modelscorewhereRAMreleased
646MistralHermes-CodePro-7B-v1 beowolx39±1local~6.1GB2023-12
647Llama-3-Instruct-8B-SimPO-v0.2 Princeton NLP39±1local~6.6GB2024-07
648Llama 3.2 3B Instruct Meta38±1api $0.12/M +local~3.4GB2024-09
649Qwen-14B-Chat Alibaba38±1local~11GB2023-09
650Falcon-180B-Chat TII38±2local~115GB2023-09
more from Meta
24Muse Spark 1.3 Meta77±1api $2.00/M2026-09
40Muse Spark 1.2 Meta75±1api $2.00/M2026-08
43Muse Spark 1.1 Meta75±1api $2.00/M2026-07

see all LLMs · find the ones that fit your machine

$ what it's good at

agent tasks▁▁▃▆█▆▅▅▅▄▃▁ #272 of 300 · 2.0% of its score · 3 tests

Beats Magistral Small 1.0 (2506), Bielik 11B v2.3 Instruct, Mixtral 8x22B Instruct v0.1 here, though they rank higher overall. Trails Granite 4.0 H 350M, Qwen3 0.6B, Qwen2.5 1.5B Instruct here, though they rank lower overall.

testresultvs the others in its rundifficultycountswhy
MASSIVE-Agents AST (10k, zero-shot)4.3% ✎█····███·███ beats 1/6hard · avg 16%▇ 0.8%sharp · too hard for it
BFCL V4 (December 2025 archive)21.9% ✎▁▁▁▂██▅▅▆▂▂▂ beats 8/77hard · avg 32%▇ 0.7%typical
BFCL v345.9% ✎▁▁▁▁▁▃▅█▅▄▁▁ beats 25/114medium · avg 52%▇ 0.5%blunt · at its level
coding▁▁▁▃▆██▇▆▄▃▁ #372 of 458 · 9.2% of its score · 7 tests

Beats LFM2.5 8B-A1B, DeepSeek R1 Distill Qwen 7B, Qwen3.5 2B here, though they rank higher overall. Trails Qwen2.5 1.5B Instruct, WizardLM 13B V1.2 here, though they rank lower overall.

testresultvs the others in its rundifficultycountswhy
Aider code editing (133 Exercism Python exercises, two attempts)26.3% ✎▁·▃▂▄▁▆▄█▆▂▂ beats 1/62medium · avg 54%▇ 3.5%at its level
HumanEval45.1% ✎▄·▄█·▄·▄···█ beats 4/7easy · avg 81%▇ 3.1%sharp · at its level
MBPP40.2% ✎▄···▄·▄▄█··█ beats 4/7medium · avg 64%▇ 1.5%at its level
CodeReviewQA: solution identification (hard), invariant accuracy7.6% ✎▃··▁▃▂▃▂█▃▄▂ beats 7/47medium · avg 54%▇ 0.4%sharp · noisy
CodeReviewQA: automated code refinement, exact match rate25.9% ✎▁▁▁▁▁▁▂▂▃█▅▄ beats 16/47hard · avg 33%▇ 0.4%typical
CodeReviewQA: change localisation (hard), invariant accuracy0.4% ✎█▂▁▁▂▂▂▃▅▅▅▇ beats 6/47medium · avg 36%▇ 0.3%sharp · noisy
CodeReviewQA: change type recognition, invariant accuracy78.8% ✎▁▁▁······▁▃█ beats 44/47easy · avg 67%▇ 0.0%blunt · at its level · noisy
chat & writing▁▁▁▃▅▄▅█▆▆▇▃ #334 of 393 · 6.2% of its score · 4 tests

Beats OpenThinker3 7B, Gemma 4 E2B, Code Llama 70B Instruct here, though they rank higher overall. Trails Granite 4.0 H 350M, LFM2 350M, Gemma 3 1B here, though they rank lower overall.

testresultvs the others in its rundifficultycountswhy
LMArena (text)1166.7 ✎▁▁▁▂▃▃▃▆▆▇█▃ beats 47/379avg 1268▇ 2.4%blunt
IFEval mean of four measures71.4% ✎█··▄▄·▄····▄ beats 0/5easy · avg 85%▇ 1.6%at its level
Arabic IFEval prompt-level (Inception)19.1% ✎▁·▁▁▂▁▄▄▆▄█▄ beats 10/38medium · avg 49%▇ 1.3%sharp
IFEval (aggregation unspecified)51.0% ✎▄····█·▄··██ beats 5/7easy · avg 68%▇ 0.9%at its level
math▁▂▃▅▆▅▄█▄▁▁▁ #344 of 398 · 19% of its score · 7 tests

Beats DBRX Instruct, Command R Plus (04-2024), Mistral NeMo Instruct (2407) here, though they rank higher overall. Trails Qwen3 0.6B, Gemma 7B (base), Qwen2.5 1.5B Instruct here, though they rank lower overall.

testresultvs the others in its rundifficultycountswhy
GSM8K +175.6% ✎█·█▄▄█···█·▄ beats 0/10very easy · avg 91%▇▇▇ 7.5%sharp · at its level · several runs
MGSM44.6% ✎▅▃▃█·▅···▃·▃ beats 1/10easy · avg 71%▇ 2.8%at its level
MATH-500 +241.2% ✎▃····▃·█···▃ beats 0/5easy · avg 81%▇ 2.6%at its level · several runs
MATH (subset unspecified)31.4% ✎▄···▄▄▄▄··█▄ beats 6/7medium · avg 53%▇ 2.5%at its level
AIME 2024 +16.7% ✎▁▁▁▃▁▃▁▅███▃ beats 6/85medium · avg 53%▇ 1.8%sharp · several runs
MATH46.7% ✎█▄██···▄▄▄·▄ beats 3/10medium · avg 64%▇ 1.7%at its level
GSM8K loose substring accuracy75.2% ✎█·█··███···█ beats 1/5easy · avg 84%▇ 0.1%blunt
other tests26 not in a task · 63% of its score
testresultvs the others in its rundifficultycountswhy
ARC Challenge +176.1% ✎▃▂▂█▂···▄··▂ beats 0/10very easy · avg 91%▇▇▇ 8.2%sharp · at its level · several runs
MMLU +261.8% ✎█▄█▄▄▄·▄▄··▄ beats 1/10easy · avg 77%▇▇▇ 7.7%at its level · several runs
BIG-Bench Hard +155.4% ✎██▄·█▄·▄▄··▄ beats 2/10medium · avg 67%▇▇▇ 7.6%sharp · several runs
OpenBookQA +172.6% ✎▃··▄█▂·▂···▄ beats 0/10easy · avg 87%▇▇ 5.0%at its level · several runs
MMLU-Pro +239.2% ✎█▄▄·█▄▄▄·▄·▄ beats 2/10medium · avg 47%▇▇ 4.9%several runs
WinoGrande +153.2% ✎▃····▃▃█▃·▅▅ beats 0/10easy · avg 84%▇▇ 4.0%at its level · several runs
HellaSwag +177.2% ✎▃▅█·▃▃▃···▃▃ beats 6/10easy · avg 87%▇ 2.9%at its level · several runs
GPQA (subset unspecified) +124.3% ✎█··▄▂··▂▂··▄ beats 2/10hard · avg 39%▇ 2.9%sharp · too hard for it · several runs
18 more tests
testresultvs the others in its rundifficultycountswhy
GPQA Diamond +230.6% ✎██····███··█ beats 4/5hard · avg 49%▇ 2.7%sharp · several runs
BoolQ +171.4% ✎▃··▃··██▃·▃▃ beats 1/10easy · avg 91%▇ 2.7%at its level · several runs
Multilingual MMLU (Microsoft)48.1% ✎█▄▄██··█···▄ beats 1/10easy · avg 65%▇ 2.0%at its level
CPTUB average (SpeakLeash, 0-5 judge scale)40.0% ✎▁▂▃▂▃█▃▆▅▄▂▃ beats 8/95easy · avg 66%▇ 1.8%at its level
Social IQA +168.3% ✎▃·▃▅▅█·▃···▃ beats 0/10easy · avg 78%▇ 1.4%blunt · at its level · several runs
Polish EQ-Bench v2 (SpeakLeash)46.5% ✎▁··▁▁▂▂▆▅▆█▆ beats 22/70medium · avg 64%▇ 1.3%at its level
Multi-IF three-turn average50.9% ✎▄··█▄··▄···▄ beats 0/5easy · avg 71%▇ 1.2%at its level
IFBench (aggregation unspecified)20.8% ✎█··███··█··█ beats 1/5hard · avg 33%▇ 1.0%typical
BRIDGE clinical text, zero-shot average (87 tasks)22.9% ✎▁▁▁▁▁▃▂▃▄▆█▃ beats 16/97hard · avg 32%▇ 0.9%blunt
ZebraLogic (ZeroEval, 1,000 puzzles)7.4% ✎▁▁▃▆██▁▁·▁▁▁ beats 4/57hard · avg 18%▇ 0.9%sharp · too hard for it
MMMLU seven languages (Liquid)47.9% ✎▄·█····█···▄ beats 0/5medium · avg 60%▇ 0.9%at its level
MGSM six languages, loose (Liquid)61.7% ✎▄·▄·█···▄··▄ beats 0/5easy · avg 82%▇ 0.8%at its level
AraGen-v3 3C3H score (Inception, Claude Sonnet 4 judge)15.0% ✎▁▁·▁▂▂▄█▄▅▄▄ beats 10/66hard · avg 31%▇ 0.6%typical
Open PT-LLM Leaderboard average (9 Portuguese tasks)64.8% ✎▃▁▁▁▁▁▁▁▂█▅▁ beats 112/276medium · avg 73%▇ 0.5%at its level · noisy
TruthfulQA MC259.2% ✎▃·█▂▂▄····▂▂ beats 0/10easy · avg 69%▇ 0.4%blunt · at its level
EuroEval Life in the UK (English knowledge MC), test split, 5-shot, accuracy77.0% ✎▁▁▁▁▁▁▁▂▄█▄▃ beats 56/146easy · avg 83%▇ 0.4%at its level · noisy
PIQA68.2% ✎▃·▅·▃█·▃▅··▃ beats 0/10easy · avg 83%▇ 0.4%blunt · at its level
GSM-Plus38.7% ✎█·····███·██ beats 0/5easy · avg 74%▇ 0.2%blunt · at its level

$ where to run it

Through an API

api · Cloudflare · $0.12/M · $0.051 in / $0.335 out per 1M tokens · 131k context · developers.cloudflare.com ↗ · price sourceWorkers 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.

llama.cpp · recommended

version

  1. Install llama.cpp · Homebrew and winget installs name the commands llama-cli and llama-server instead.
    macOS / Linux
    curl -LsSf https://llama.app/install.sh | sh
    Windows
    irm https://llama.app/install.ps1 | iex

    other ways to install ↗

  2. Run Llama 3.2 3B Instruct · needs ~2 GB · from huggingface.co/lmstudio-community
    chat in the terminal
    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)
    llama serve -hf lmstudio-community/Llama-3.2-3B-Instruct-GGUF:Q4_K_M

LM Studio

version

  1. Install LM Studio · Open the app once before using the lms command.

    download it (macOS / Windows / Linux) ↗

  2. Run Llama 3.2 3B Instruct · needs ~2 GB · from huggingface.co/lmstudio-community

    open in LM Studio ↗

  3. Or download it in the terminal
    lms get https://huggingface.co/lmstudio-community/Llama-3.2-3B-Instruct-GGUF@Q4_K_M -y
  4. Then run lms ls to see its name, and lms chat followed by that name

Ollama

version

  1. Install Ollama

    download it (macOS / Windows) ↗

    Linux
    curl -fsSL https://ollama.com/install.sh | sh
  2. Run Llama 3.2 3B Instruct · needs ~2.43 GB · from ollama.com
    chat in the terminal
    ollama run llama3.2:3b

MLX (Mac)

version

  1. Install MLX (Mac) · Apple silicon Macs only, text only.
    with uv
    uv tool install mlx-lm
    or
    pip install mlx-lm
  2. Run Llama 3.2 3B Instruct · needs ~1.8 GB, Mac only · from huggingface.co/mlx-community-staging
    chat in the terminal
    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
    mlx_lm.server --model mlx-community-staging/Llama-3.2-3B-Instruct-mlx-4Bit

Weights and documentation: huggingface.co/meta-llama ↗

$ how the score was worked out · microsoft/Phi-4-mini-instruct… 36%, IBM Granite 3.0 Language Mode… 31%, LFM2 Technical Report — Liqui… 11% and 14 more

Each result implies a score on its own, after adjusting for how hard its test is and who ran it. The score is their weighted average, with a small pull toward the middle for thinly tested models. Each test's weight is its counts column above; here it's added up by who ran the tests. How it works

sourceresultsshare of the scorelicence
microsoft/Phi-4-mini-instruct model card16▇▇▇▇▇▇▇▇▇▇▇ 36%MIT
IBM Granite 3.0 Language Models technical report (GitHub)15▇▇▇▇▇▇▇▇▇ 31%Apache-2.0
LFM2 Technical Report — Liquid AI own SmolLM3 evaluation11▇▇▇ 11%CC BY 4.0
Aider code editing benchmark runs (old edit leaderboard)1▇ 3%Apache-2.0
microsoft/Phi-4-mini-reasoning model card3▇ 3%MIT
Microsoft — Phi-4-Mini Technical Report3▇ 2%CC BY 4.0
LMArena leaderboard dataset1▇ 2%CC BY 4.0
Inception Arabic/Hindi leaderboards (AraGen-v3, HindiGen-v1, Arabic IFEval)2▇ 2%Apache-2.0
SpeakLeash CPTUB (Complex Polish Text Understanding Benchmark)1▇ 2%MIT
SpeakLeash Polish EQ-Bench leaderboard1▇ 1%Apache-2.0
Berkeley Function Calling Leaderboard result archive2▇ 1%Apache-2.0
CodeReviewQA: The Code Review Comprehension Assessment for Large Language Models (arXiv:2503.16167v2)4▇ 1%CC-BY-4.0
BRIDGE Medical Leaderboard (Harvard Medical School / Brigham and Women's YLab)1▇ 1%MIT
WildEval ZeroEval published results1▇ 1%Apache-2.0
MASSIVE-Agents — Kulkarni et al. (2025)1▇ 1%CC BY 4.0
Open Portuguese LLM Leaderboard (archived)1▇ 0%Apache-2.0
EuroEval leaderboards (results collection), English generative leaderboard1▇ 0%MIT
prior · pull toward the middle▇ 0%
weighted average= 38.4
$ published speed measurements · 1

Reference results for the hardware, engine and settings below. They aren't verified speeds of the downloads above or predictions for your machine.

Apple M4 Pro (12-core CPU / 16-core GPU, 48GB)GGUF Q4_K_M (Ollama registry default)115.2 tok/sOllama v0.21.2 (llama.cpp, Metal); macOS 15.4 · throughput = eval_count / eval_duration (Ollama API; decode phase only); temperature 0; one warm-up per batch; one model resident at a time; Table IV model-level value over the MMLU (~15 output tokens), summarization and code-generation (~200 output tokens) tasks; aggregation across tasks not further specified

Model-level figure from Table IV; per-task rates differ substantially (paper text).

source ↗
$ quantization evidence · 1 measurements

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.

precisiontestfull → quantizedkept
ONNX INT4 AWQ (NVIDIA ModelOpt-Windows v0.19.0, quantizing the FP16 ONNX model)MMLU 5-shot (ModelOpt-Windows accuracy_benchmark; onnxruntime-genai-directml 0.4, transformers 4.44)0.608 → 0.57794.9%source ↗

$ improve this page

Update model informationAdd a place to run itAdd a test resultPropose a benchmarkCorrect a benchmark scoreEdit data on GitHub

Use a short GitHub form or edit the data through a pull request. Every number needs a source; here's which ones we can use.

data updated · sources and licences · json