whatmodel.sh / Nemotron 3 Nano 4B

Nemotron 3 Nano 4B

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

Nemotron 3 Nano 4B is an open-weights language model from NVIDIA, released March 2026. Its published size is 3.97B parameters and a 262k-token context window. Licence: nvidia-open-model.

54.3±0.9#281 of 719 LLMsfrom 26 benchmark results across 2 sources

$ how it compares

The LLMs ranked either side of Nemotron 3 Nano 4B, out of 828 with a score, then more from NVIDIA.

#modelscorewhereRAMreleased
333Llama 3.3 70B Instruct Meta54±1api $0.72/M +local~24GB2024-12
334Llama 4 Scout (109B-A17B) Meta54±1api $0.15/M +local~36GB2025-04
335Nemotron 3 Nano 4B NVIDIA54±1local~3.8GB2026-03
336Granite 4.1 8B IBM54±1api $0.11/M +local~7GB2026-04
337Pixtral Large (2411) Mistral54±3local~51GB2024-11
more from NVIDIA
98Nemotron 3 Ultra 550B-A55B NVIDIA69±1api $1.35/M +local~89GB2026-06
155Nemotron Cascade 2 30B-A3B NVIDIA66±1local~20GB2026-03
168Nemotron 3 Super 120B-A12B NVIDIA65±1api $0.28/M +local~73GB2026-03

see all LLMs · find the ones that fit your machine

$ what it's good at

agent tasks▁▁▃▆█▆▅▅▅▄▃▁ #242 of 300 · 21% of its score · 7 tests

Beats DeepSeek R1 0528 Qwen3 8B, Sky-T1 32B Preview, Magistral Small 1.0 (2506) here, though they rank higher overall. Trails Granite 4.0 1B, SmolLM3 3B, Phi-3.5-mini-instruct here, though they rank lower overall.

testresultvs the others in its rundifficultycountswhy
τ²-bench Telecom lab33.0% ✎▁·▅█▆▅▂▃▃▂▃▁ beats 27/89hard · avg 19%▇▇▇▇ 10%sharp · at its level
BFCL V4 (December 2025 archive)43.7% ✎▄·█·▄▄▄▄·▄·▄ beats 2/8medium · avg 43%▇ 2.7%at its level
BrowseComp-ZH3.3% ✎▄▄·▄▄▄·█···█ beats 0/8very hard · avg 13%▇ 2.4%typical
SWE-bench Verified3.0% ✎▄··▄▄█··▄·▄█ beats 2/8very hard · avg 12%▇ 2.3%sharp · noisy
SWE-bench Pro (split unspecified)0.1% ✎█▂···▂···▄·▂ beats 1/8very hard · avg 4.3%▇ 1.6%sharp · too hard for it · noisy
τ³-bench Banking1.2% ✎▃▂·▅█▂▂▃▂▅▂▃ beats 0/20very hard · avg 3.7%▇ 1.3%blunt · too hard for it
BFCL v3 lab +161.1% ✎▁▁▁▁▁▃▅█▆▅▄▁ beats 102/170medium · avg 58%▇ 0.8%blunt · at its level · small run
coding▁▁▁▃▆██▇▆▄▃▁ #275 of 458 · 19% of its score · 6 tests

Beats Qwen3-VL 32B Instruct, Solar Open 100B (102B-A12B), GLM-4.7-Flash here, though they rank higher overall. Trails DeepSeek LLM 67B Chat, Yi 1.5 9B Chat, Mistral NeMo Instruct (2407) here, though they rank lower overall.

testresultvs the others in its rundifficultycountswhy
LiveCodeBench v6 (window unspecified)50.7% ✎▃····█·▃▅▃·▃ beats 4/8medium · avg 39%▇▇ 5.2%at its level
LiveCodeBench Pro 2025Q2 (Easy)51.6% ✎▄···█▄··▄█▄▄ beats 5/8hard · avg 33%▇▇ 4.2%at its level
LiveCodeBench Pro 2025Q2 (Medium)5.3% ✎█···▂··▄▂··▂ beats 5/8very hard · avg 2.0%▇ 3.1%sharp · too hard for it · noisy
SWE-bench Verified3.0% ✎▄··▄▄█··▄·▄█ beats 2/8very hard · avg 12%▇ 2.3%sharp · noisy
LiveCodeBench (version unspecified) lab +151.8% ✎▁···▁▁▃▅▅█▄▅ beats 18/37medium · avg 44%▇ 2.0%at its level · small run
SWE-bench Pro (split unspecified)0.1% ✎█▂···▂···▄·▂ beats 1/8very hard · avg 4.3%▇ 1.6%sharp · too hard for it · noisy
chat & writing▁▁▁▃▅▄▅█▆▆▇▃ #160 of 393 * · 0.2% of its score · 2 tests

Beats MiniMax-M2, MiniCPM5 2B, Granite 4.2 30B here, though they rank higher overall. Trails LFM2.5 1.2B Thinking, LFM2.5 8B-A1B, LFM2.5 2.6B here, though they rank lower overall.

testresultvs the others in its rundifficultycountswhy
IFEval (aggregation unspecified)88.0% ✎▄▄····▄··███ beats 4/8easy · avg 77%▇ 0.2%noisy
IFEval prompt-level (strictness unspecified) lab87.9% ✎no other model yet—not counted yet: no other model in the same run to compare it with
math▁▂▃▅▆▅▄█▄▁▁▁ #187 of 398 · 38% of its score · 4 tests

Beats MiniMax-M2.1, DeepSeek V3 0324, Qwen3 235B-A22B (Apr 2025, hybrid) here, though they rank higher overall. Trails Phi-4-mini-reasoning 3.8B, DeepSeek R1 Distill Qwen 7B, OpenThinker3 7B here, though they rank lower overall.

testresultvs the others in its rundifficultycountswhy
AIME 2025 +256.3% ✎█▄▄▄▄····█·▄ beats 5/8medium · avg 60%▇▇▇▇▇▇ 16%sharp · at its level · several runs
HMMT February 202651.3% ✎▅··▃▃·▃··▃·█ beats 5/8hard · avg 30%▇▇▇▇ 10%sharp · at its level
AIME 202662.1% ✎▄·▄█·▄▄···█▄ beats 5/8medium · avg 48%▇▇▇ 6.6%sharp · at its level
MATH-500 +291.6% ✎██▄▄▄··▄···▄ beats 4/8very easy · avg 89%▇▇▇ 6.5%maxed out · several runs
web search█·█▄██·▄▄·▄▄ #13 of 13 * · 2.4% of its score · 1 test

Trails Gemma 4 E2B, Qwen3.5 2B, LFM2.5 8B-A1B here, though they rank lower overall.

testresultvs the others in its rundifficultycountswhy
BrowseComp-ZH3.3% ✎▄▄·▄▄▄·█···█ beats 0/8very hard · avg 13%▇ 2.4%typical
other tests8 not in a task · 26% of its score
testresultvs the others in its rundifficultycountswhy
GPQA (subset unspecified) lab +153.2% ✎▂▁▂▂▄▃▆▃▅█▄▃ beats 54/104hard · avg 48%▇▇▇ 7.4%sharp · at its level
MMLU-Pro65.7% ✎▂··▂█·▂▂···▂ beats 4/8medium · avg 59%▇▇ 4.1%at its level
MMLU-Redux79.8% ✎▂···█··▂▂··▂ beats 2/8easy · avg 77%▇ 3.6%typical
OJBench20.0% ✎▃····▃▅··█▃▃ beats 5/8very hard · avg 12%▇ 2.9%typical
SuperGPQA37.8% ✎▃·▃·▃▅█····▃ beats 3/8hard · avg 32%▇ 2.5%at its level
LongBench v232.0% ✎▃······▅▅▃·█ beats 3/8very hard · avg 32%▇ 2.3%typical
IFBench (aggregation unspecified)58.3% ✎▅····▃▃·█·▃▃ beats 4/8medium · avg 43%▇ 1.7%at its level
Multi-IF (aggregation unspecified)65.9% ✎▄▄···▄·▄·▄██ beats 3/8medium · avg 60%▇ 1.2%blunt · at its level

$ where to run it

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 Nemotron 3 Nano 4B · needs ~2.8 GB · from huggingface.co/lmstudio-community
    chat in the terminal
    llama cli -hf lmstudio-community/NVIDIA-Nemotron-3-Nano-4B-GGUF:Q4_K_M
  3. Or start a local server and open the address it prints (http://localhost:9931)
    llama serve -hf lmstudio-community/NVIDIA-Nemotron-3-Nano-4B-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 Nemotron 3 Nano 4B · needs ~2.8 GB · from huggingface.co/lmstudio-community

    open in LM Studio ↗

  3. Or download it in the terminal
    lms get https://huggingface.co/lmstudio-community/NVIDIA-Nemotron-3-Nano-4B-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 Nemotron 3 Nano 4B · needs ~2.88 GB · from ollama.com
    chat in the terminal
    ollama run nemotron-3-nano:4b

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 Nemotron 3 Nano 4B · needs ~2.2 GB, Mac only · from huggingface.co/mlx-community
    chat in the terminal
    mlx_lm.chat --model mlx-community/NVIDIA-Nemotron-3-Nano-4B-4bit
  3. Or start an OpenAI-compatible server at http://localhost:8080/v1
    mlx_lm.server --model mlx-community/NVIDIA-Nemotron-3-Nano-4B-4bit

Official weights and documentation: huggingface.co/nvidia ↗

$ how the score was worked out · openbmb/MiniCPM5-2B model card 68%, NVIDIA vendor-reported benchm… 31%

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
openbmb/MiniCPM5-2B model card20▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇ 68%Apache-2.0
NVIDIA vendor-reported benchmark scores for its own Nemotron models (Hugging Face model cards) · lab-reported6▇▇▇▇▇▇▇▇▇ 31%Factual vendor-reported score (not licensed content)
prior · pull toward the middle▇ 0%
weighted average= 54.3

not counted (6)

nvidia/NVIDIA-Nemotron-3-Nano-4B-BF16 model cardAIME 2025 [correct]78.5%the only model in this source's run of the test, so there is nothing to compare it with there; another run of the test counts
nvidia/NVIDIA-Nemotron-3-Nano-4B-BF16 model cardMATH-500 [correct]95.4%the only model in this source's run of the test, so there is nothing to compare it with there; another run of the test counts
nvidia/NVIDIA-Nemotron-3-Nano-4B-BF16 model cardGPQA (subset unspecified) [correct]53.2%the only model in this source's run of the test, so there is nothing to compare it with there; another run of the test counts
nvidia/NVIDIA-Nemotron-3-Nano-4B-BF16 model cardLiveCodeBench (version unspecified) [correct]51.8%the only model in this source's run of the test, so there is nothing to compare it with there; another run of the test counts
nvidia/NVIDIA-Nemotron-3-Nano-4B-BF16 model cardBFCL v3 [correct]61.1%the only model in this source's run of the test, so there is nothing to compare it with there; another run of the test counts
nvidia/NVIDIA-Nemotron-3-Nano-4B-BF16 model cardIFEval prompt-level (strictness unspecified) [correct]87.9%fewer than three models have this test yet, or this source ran it for this model alone
$ published speed measurements · 5

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

NVIDIA DGX Spark (GB10)Q8_052.85 tok/sllama.cpp 1bbec6a75 (8382), CUDA · llama-bench; tg32; 32 generated tokens; context depth 0; n_ubatch=2048; flash attention=1; one sequence

Published token-generation mean; not prompt processing or aggregate batch throughput. Hardware core count and installed RAM are not inferred.

source ↗
NVIDIA DGX Spark (GB10)Q8_052.32 tok/sllama.cpp 1bbec6a75 (8382), CUDA · llama-bench; tg32 @ d4096; 32 generated tokens; context depth 4096; n_ubatch=2048; flash attention=1; one sequence

Published token-generation mean; not prompt processing or aggregate batch throughput. Hardware core count and installed RAM are not inferred.

source ↗
NVIDIA DGX Spark (GB10)Q8_051.27 tok/sllama.cpp 1bbec6a75 (8382), CUDA · llama-bench; tg32 @ d8192; 32 generated tokens; context depth 8192; n_ubatch=2048; flash attention=1; one sequence

Published token-generation mean; not prompt processing or aggregate batch throughput. Hardware core count and installed RAM are not inferred.

source ↗
NVIDIA DGX Spark (GB10)Q8_049.71 tok/sllama.cpp 1bbec6a75 (8382), CUDA · llama-bench; tg32 @ d16384; 32 generated tokens; context depth 16384; n_ubatch=2048; flash attention=1; one sequence

Published token-generation mean; not prompt processing or aggregate batch throughput. Hardware core count and installed RAM are not inferred.

source ↗
NVIDIA DGX Spark (GB10)Q8_046.71 tok/sllama.cpp 1bbec6a75 (8382), CUDA · llama-bench; tg32 @ d32768; 32 generated tokens; context depth 32768; n_ubatch=2048; flash attention=1; one sequence

Published token-generation mean; not prompt processing or aggregate batch throughput. Hardware core count and installed RAM are not inferred.

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