whatmodel.sh / Nemotron 3 Nano 30B-A3B

Nemotron 3 Nano 30B-A3B

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

Nemotron 3 Nano 30B-A3B is an open-weights language model from NVIDIA, released December 2025. Its published size is 31.6B parameters (3.6B active per token) and a 262k-token context window. Licence: nvidia-open-model.

57.5±0.9#227 of 719 LLMsfrom 22 benchmark results across 9 sources

$ how it compares

The LLMs ranked either side of Nemotron 3 Nano 30B-A3B, out of 828 with a score, then more from NVIDIA.

#modelscorewhereRAMreleased
269Nova 2 Lite Amazon58±2api $0.94/M2025-10
270Llama 3.3 Nemotron Super 49B v1.5 NVIDIA58±1local~18GB2025-07
271Nemotron 3 Nano 30B-A3B NVIDIA58±1api $0.087/M +local~20GB2025-12
272Ministral 3 14B Reasoning 2512 Mistral58±1local~10GB2025-12
273GPT-4o (2024-08-06) OpenAI57±1api $4.38/M2024-08
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▁▁▃▆█▆▅▅▅▄▃▁ #175 of 300 * · 17% of its score · 5 tests

Beats o1 (2024-12-17), o3-mini, DeepSeek R1 0528 here, though they rank higher overall. Trails SmolLM3 3B, Jamba 1.5 Mini, Jamba 1.5 Large here, though they rank lower overall.

testresultvs the others in its rundifficultycountswhy
τ²-bench Telecom lab42.2% ✎▁·▅█▆▅▂▃▃▂▃▁ beats 35/89hard · avg 19%▇▇▇▇▇ 12%sharp · at its level
Terminal-Bench Hard lab8.5% ✎▂▅▆▂▂▆▂▂█▅▅▂ beats 11/27very hard · avg 4.9%▇▇ 4.5%sharp · small run
BFCL V4 (December 2025 archive) lab53.8% ✎▁▁▁▂▇▆▆▆█▂▄▂ beats 81/111hard · avg 32%—not counted yet: no other model in the same run to compare it with
Humanity's Last Exam with tools lab15.5% ✎▃▂▂▁▄▂▃▆█▇▂▆ beats 0/54hard · avg 16%—not counted yet: no other model in the same run to compare it with
τ²-bench aggregate lab49.0% ✎▁▁▁·▂▃▄▃▃▅█▂ beats 6/37medium · avg 40%—not counted yet: no other model in the same run to compare it with
coding▁▁▁▃▆██▇▆▄▃▁ #250 of 458 · 10% of its score · 5 tests

Beats Qwen3-VL 32B Instruct, Solar Open 100B (102B-A12B), GLM-4.7-Flash here, though they rank higher overall. Trails QwQ-32B-Preview, InternLM2.5 20B Chat, Mistral Small Instruct 22B (2409) here, though they rank lower overall.

testresultvs the others in its rundifficultycountswhy
LiveCodeBench Pro 2025Q2 (Easy) lab54.5% ✎▂··▃▂▃▆▅█·▃▂ beats 8/20medium · avg 45%▇▇ 4.5%at its level
LiveCodeBench Pro 2025Q2 (Medium) lab3.5% ✎█··▄▂▃██▄·▄▆ beats 7/22very hard · avg 3.0%▇ 3.7%sharp · too hard for it
SciCode (level unspecified) lab33.3% ✎▁····▃▄▄▄█▆▂ beats 15/34hard · avg 19%▇ 2.1%at its level · small run
LiveCodeBench v6 (Aug 2024–May 2025) lab68.3% ✎▂▅▅▃█▃▂····▂ beats 8/17medium · avg 45%—not counted yet: no other model in the same run to compare it with
SciCode subproblems lab33.3% ✎▃··▃▃▃·▃·█·▅ beats 0/9hard · avg 17%—not counted yet: no other model in the same run to compare it with
chat & writing▁▁▁▃▅▄▅█▆▆▇▃ #208 of 393 * · 7.5% of its score · 2 tests

Beats Granite 4.2 8B, Qwen3-VL 30B-A3B Thinking, Mistral Small 4 (119B-A6.5B) here, though they rank higher overall. Trails LFM2.5 1.2B Instruct, Granite 4.0 H Micro 3B, Gemma 3n E4B here, though they rank lower overall.

testresultvs the others in its rundifficultycountswhy
LMArena (text)1313.5 ✎▁▁▁▂▃▃▃▆▆▇█▃ beats 143/379avg 1268▇▇▇ 7.5%blunt
Arena-Hard v2 aggregate lab67.7% ✎▁▁··▄▃▁▃█▁▁▇ beats 14/22hard · avg 27%—not counted yet: no other model in the same run to compare it with
math▁▂▃▅▆▅▄█▄▁▁▁ #143 of 398 · 25% of its score · 4 tests

Beats Claude Sonnet 4.6, MiMo-V2.5 (310B-A15B), MiniMax-M2.7 here, though they rank higher overall.

testresultvs the others in its rundifficultycountswhy
AIME 2025 lab +189.1% ✎▁·▁▁▁▂▅▆█▇▅▁ beats 111/133medium · avg 63%▇▇▇ 8.5%sharp
AIME 2026 lab89.9% ✎▁▁▁▂▁▂▆█▆▂·▁ beats 22/49medium · avg 64%▇▇▇ 7.0%sharp
IMO-AnswerBench lab70.4% ✎▁▁·▁▁▁▄▅█▄▂▃ beats 10/42medium · avg 42%▇▇ 6.0%at its level
HMMT February 2025 lab84.6% ✎▁▁▁▃▁▅▅▃██▁▄ beats 15/31easy · avg 69%▇▇ 4.0%small run
web searchnot ranked on this task yet · 1 test
testresultvs the others in its rundifficultycountswhy
Humanity's Last Exam with tools lab15.5% ✎▃▂▂▁▄▂▃▆█▇▂▆ beats 0/54hard · avg 16%—not counted yet: no other model in the same run to compare it with
other tests15 not in a task · 40% of its score
testresultvs the others in its rundifficultycountswhy
Medmarks-V mean win rate (19 verifiable medical benchmarks, 71 configurations)0.5114 ✎▁▄▃▃▇▄█▃▅▁▁▁ beats 19/30avg 0.423▇▇▇▇▇ 11%sharp
Medmarks-V weighted mean win rate (61 models, 71 configurations)51.1% ✎▂▃▅▂▆▃█▅█▅▃▆ beats 19/34medium · avg 42%▇▇ 5.7%blunt · at its level
EuroEval Life in the UK (English knowledge MC), validation split, 5-shot, accuracy93.0% ✎▂·▂▂·▂·▃▃·█▃ beats 9/14easy · avg 80%▇▇ 4.5%typical
MMLU-ProX lab +159.5% ✎▁·▁▁▃▂▁▃▂█▅▄ beats 3/35medium · avg 54%▇▇ 4.2%at its level · small run · several runs
CPTUB average (SpeakLeash, 0-5 judge scale)64.9% ✎▁▂▂▂▃█▃▆▅▄▂▃ beats 43/95easy · avg 66%▇▇ 4.2%typical
Polish EQ-Bench v2 (SpeakLeash)45.8% ✎▁··▁▁▂▂▆▅▆█▆ beats 17/70medium · avg 64%▇ 3.1%typical
EuroEval HellaSwag (English), validation split, 5-shot generative MC, accuracy69.2% ✎▂··▂▂·▃▆▆█·▄ beats 3/15medium · avg 63%▇ 3.0%at its level
MultiChallenge lab +138.5% ✎▁·▁·▁▃▆▅▃▁█▅ beats 3/29hard · avg 27%▇ 2.2%at its level · small run
7 more tests
testresultvs the others in its rundifficultycountswhy
CL-bench lab12.0% ✎█··········▄ beats 0/2very hard · avg 7.1%▇ 1.2%blunt · small run
miniF2F (pass@1) lab50.0% ✎▃·█▅·····▅·▃ beats 7/8very hard · avg 14%▇ 0.3%blunt · noisy · small run
Vectara hallucination leaderboard (HHEM-2.3, factual consistency rate)90.4% ✎▁▂▁▂▅█▂▃▄▃▁▁ beats 48/98very easy · avg 90%▇ 0.0%blunt
MMLU-Pro lab78.3% ✎▁▁·▁▁▁▃▂▄▅█▁ beats 116/188medium · avg 57%—not counted yet: no other model in the same run to compare it with
GPQA (subset unspecified) lab73.0% ✎▂▁▂▂▄▃▆▃▅█▄▃ beats 81/104very hard · avg 33%—not counted yet: no other model in the same run to compare it with
Humanity's Last Exam lab10.6% ✎▁▁▁▃▆▆▇██▆▁▃ beats 19/104very hard · avg 3.8%—not counted yet: no other model in the same run to compare it with
IFBench prompt-level lab71.5% ✎█▄▄··▄·▄···▄ beats 0/6—not counted yet: no other model in the same run to compare it with

$ where to run it

Through an API

api · Deep Infra · $0.088/M · $0.05 in / $0.2 out per 1M tokens · 262k context · deepinfra.com ↗ · price sourceStandard base token rate from models.dev. Provider-specific tiers, long-context surcharges, caching and promotions may differ.

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 30B-A3B · needs ~22.4 GB · from huggingface.co/ggml-org
    chat in the terminal
    llama cli -hf ggml-org/NVIDIA-Nemotron-3-Nano-30B-A3B-GGUF:Q4_K_M
  3. Or start a local server and open the address it prints (http://localhost:9931)
    llama serve -hf ggml-org/NVIDIA-Nemotron-3-Nano-30B-A3B-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 30B-A3B · needs ~22.4 GB · from huggingface.co/ggml-org

    open in LM Studio ↗

  3. Or download it in the terminal
    lms get https://huggingface.co/ggml-org/NVIDIA-Nemotron-3-Nano-30B-A3B-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 30B-A3B · needs ~19.46 GB · from ollama.com
    chat in the terminal
    ollama run nemotron-3-nano:30b

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 30B-A3B · needs ~17.8 GB, Mac only · from huggingface.co/lmstudio-community
    chat in the terminal
    mlx_lm.chat --model lmstudio-community/NVIDIA-Nemotron-3-Nano-30B-A3B-MLX-4bit
  3. Or start an OpenAI-compatible server at http://localhost:8080/v1
    mlx_lm.server --model lmstudio-community/NVIDIA-Nemotron-3-Nano-30B-A3B-MLX-4bit

Official weights and documentation: huggingface.co/nvidia ↗

$ how the score was worked out · NVIDIA — Nemotron-Cascade 2 t… 41%, Medmarks: A Comprehensive Ope… 17%, NVIDIA vendor-reported benchm… 17% and 6 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
NVIDIA — Nemotron-Cascade 2 technical report · lab-reported10▇▇▇▇▇▇▇▇▇▇▇▇ 41%CC BY 4.0
Medmarks: A Comprehensive Open-Source LLM Benchmark Suite for Medical Tasks — authors' own runs of 61 models2▇▇▇▇▇ 17%CC BY 4.0
NVIDIA vendor-reported benchmark scores for its own Nemotron models (Hugging Face model cards) · lab-reported2▇▇▇▇▇ 17%Factual vendor-reported score (not licensed content)
EuroEval leaderboards (results collection), English generative leaderboard2▇▇ 8%MIT
LMArena leaderboard dataset1▇▇ 7%CC BY 4.0
SpeakLeash CPTUB (Complex Polish Text Understanding Benchmark)1▇ 4%MIT
SpeakLeash Polish EQ-Bench leaderboard1▇ 3%Apache-2.0
NVIDIA — Nemotron 3 Nano technical report · lab-reported2▇ 2%CC BY 4.0
Vectara hallucination leaderboard (vectara/hallucination-leaderboard repository)1▇ 0%Apache-2.0
prior · pull toward the middle▇ 0%
weighted average= 57.5

not counted (12)

nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16 model cardMMLU-Pro [correct]78.3%fewer than three models have this test yet, or this source ran it for this model alone
nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16 model cardAIME 2025 [correct]89.1%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-30B-A3B-BF16 model cardGPQA (subset unspecified) [correct]73.0%fewer than three models have this test yet, or this source ran it for this model alone
nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16 model cardSciCode subproblems [correct]33.3%fewer than three models have this test yet, or this source ran it for this model alone
nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16 model cardHumanity's Last Exam [correct]10.6%fewer than three models have this test yet, or this source ran it for this model alone
nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16 model cardτ²-bench aggregate [correct]49.0%fewer than three models have this test yet, or this source ran it for this model alone
nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16 model cardBFCL V4 (December 2025 archive) [correct]53.8%fewer than three models have this test yet, or this source ran it for this model alone
nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16 model cardIFBench prompt-level [correct]71.5%fewer than three models have this test yet, or this source ran it for this model alone
nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16 model cardArena-Hard v2 aggregate [correct]67.7%fewer than three models have this test yet, or this source ran it for this model alone
NVIDIA — Nemotron 3 Nano technical reportHumanity's Last Exam with tools [correct]15.5%fewer than three models have this test yet, or this source ran it for this model alone
NVIDIA — Nemotron 3 Nano technical reportLiveCodeBench v6 (Aug 2024–May 2025) [correct]68.3%fewer than three models have this test yet, or this source ran it for this model alone
NVIDIA — Nemotron 3 Nano technical reportMultiChallenge [correct]38.5%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
$ 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.

AMD Ryzen AI Max+ 395 (Strix Halo), Radeon 8060S iGPU; Beelink GTR9 Pro 128GBIQ4_XS75.97 tok/sllama.cpp 2016bf2 (build_number 1 as recorded), Vulkan (RADV) · llama-bench; tg128; 128 generated tokens; context depth 0; n_prompt=0 row; n_batch=2048; n_ubatch=512; flash_attn=-1; type_k=f16; type_v=f16; n_gpu_layers=-1; threads=16; one sequence; file label r5 repetitions

Author's own llama-bench CSV row (avg_ts mean over repetitions, stddev_ts retained). GPU string as recorded: Radeon 8060S Graphics (RADV STRIX_HALO). Unified memory; the guide reports ~124GiB OS-visible. Kernel/Mesa/power profile per the run folder notes. Not prompt processing or multi-request throughput. Curated in the guide's data/benchmarks.csv line 37 (2026-06-05, measured-local): Latest/int-dot practical NVIDIA 30B-class route after the Nemotron 3 release; current-model scout, not a top speed headline.

source ↗
$ quantization evidence · 6 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
NVFP4 PTQ (post-training, W4A4 FP4)GPQA-Diamond accuracy (NVIDIA evaluation settings)0.73 → 0.71698.1%source ↗
NVFP4 PTQ (post-training, W4A4 FP4)LiveCodeBench v5 accuracy (NVIDIA evaluation settings)0.721 → 0.68995.6%source ↗
NVFP4 PTQ (post-training, W4A4 FP4)AIME 2025 accuracy (NVIDIA evaluation settings)0.891 → 0.8595.4%source ↗
NVFP4 QAD (quantization-aware distillation)GPQA-Diamond accuracy (NVIDIA evaluation settings)0.73 → 0.72799.6%source ↗
NVFP4 QAD (quantization-aware distillation)AIME 2025 accuracy (NVIDIA evaluation settings)0.891 → 0.87998.7%source ↗
NVFP4 QAD (quantization-aware distillation)LiveCodeBench v5 accuracy (NVIDIA evaluation settings)0.721 → 0.68995.6%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