whatmodel.sh / Nemotron 3 Super 120B-A12B

Nemotron 3 Super 120B-A12B

+ contribute

Nemotron 3 Super 120B-A12B is an open-weights language model from NVIDIA, released March 2026. Its published size is 120.6B parameters (12.7B active per token) and a 262k-token context window. Licence: nvidia-open-model.

64.7±0.8#146 of 719 LLMsfrom 26 benchmark results across 5 sources

$ how it compares

The LLMs ranked either side of Nemotron 3 Super 120B-A12B, out of 828 with a score, then more from NVIDIA.

#	model	score	where	RAM	released
166	ChatGPT-4o (2025-03-26) OpenAI	65±2	comparison only		2025-03
167	INTELLECT-3 Prime Intellect	65±1	local	~48GB	2025-11
168	Nemotron 3 Super 120B-A12B NVIDIA	65±1	api $0.28/M +local	~73GB	2026-03
169	GLM-4.5 Z.ai	65±1	api $1.00/M +local	~58GB	2025-07
170	Solar Pro 4 Upstage	65±2	api $0.53/M		2026-08
	more from NVIDIA
98	Nemotron 3 Ultra 550B-A55B NVIDIA	69±1	api $1.35/M +local	~89GB	2026-06
155	Nemotron Cascade 2 30B-A3B NVIDIA	66±1	local	~20GB	2026-03
177	Nemotron-Labs-3 Puzzle 75B-A9B NVIDIA	64±1	local	~35GB	2026-06

see all LLMs · find the ones that fit your machine

$ what it's good at

agent tasks▁▁▃▆█▆▅▅▅▄▃▁ #128 of 300 · 24% of its score · 10 tests

Beats o3, Gemini 2.5 Pro, Gemini 3.1 Flash-Lite here, though they rank higher overall. Trails SmolLM3 3B, Jamba 1.5 Large, MiniCPM5 2B here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
τ²-bench Telecom lab	64.4% ✎	▁·▅█▆▅▂▃▃▂▃▁ beats 51/89	hard · avg 19%	▇▇▇▇ 10%	sharp · at its level
Terminal-Bench Hard lab	25.8% ✎	▂▅▆▂▂▆▂▂█▅▅▂ beats 20/27	very hard · avg 4.9%	▇▇ 5.6%	sharp · small run
SWE-bench Verified lab	59.5% ✎	▁▁▁▁▁▁▂▄█▆▁▁ beats 53/131	hard · avg 21%	▇▇ 3.9%	sharp · at its level · small run
τ²-bench Retail lab	64.3% ✎	▁····▂▁▃▆█▃▁ beats 16/44	medium · avg 35%	▇ 2.3%	at its level · small run
τ²-bench Airline lab	56.9% ✎	▁▁·▂▁▃▂▁█▅▂▃ beats 20/42	hard · avg 29%	▇ 1.9%	at its level · small run
SWE-bench Multilingual lab	45.8% ✎	▁··▁▃▃▅▇█▃▃▁ beats 10/60	hard · avg 18%	—	not counted yet: no other model in the same run to compare it with
Humanity's Last Exam with tools lab	22.8% ✎	▂▂▃▁▄▂▃▆█▇▂▆ beats 5/54	hard · avg 16%	—	not counted yet: no other model in the same run to compare it with
BrowseComp lab	31.3% ✎	▂·▄▁▂▃▆▄█▂▃▄ beats 11/42	very hard · avg 10%	—	not counted yet: no other model in the same run to compare it with

2 more tests

test	result	vs the others in its run	difficulty	counts	why
τ²-bench aggregate lab	61.2% ✎	▁▁▁·▂▂▄▃▃▅█▂ beats 11/37	medium · avg 40%	—	not counted yet: no other model in the same run to compare it with
Terminal-Bench 2.0 lab	31.0% ✎	▂▂▅▅████▆▂▃▃ beats 6/36	very hard · avg 9.5%	—	not counted yet: no other model in the same run to compare it with

coding▁▁▁▃▆██▇▆▄▃▁ #117 of 458 · 16% of its score · 7 tests

Beats Qwen3.5 Plus, Gemini 2.5 Pro, Kimi K2 Thinking here, though they rank higher overall. Trails Llama 3.1 Nemotron Ultra 253B v1, Solar Pro 4 here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
LiveCodeBench Pro 2025Q2 (Medium) lab	23.2% ✎	█··▄▂▂██▄·▄▆ beats 20/22	very hard · avg 3.0%	▇▇ 5.6%	sharp
SWE-bench Verified lab	59.5% ✎	▁▁▁▁▁▁▂▄█▆▁▁ beats 53/131	hard · avg 21%	▇▇ 3.9%	sharp · at its level · small run
LiveCodeBench Pro 2025Q2 (Easy) lab	81.7% ✎	▂··▃▂▃▆▅█·▃▂ beats 19/20	medium · avg 45%	▇ 3.1%	typical
SciCode (level unspecified) lab	42.1% ✎	▁····▃▄▄▄█▆▂ beats 28/34	hard · avg 19%	▇ 2.0%	at its level · small run
SciCode subproblems lab +1	42.3% ✎	▃··▃▃▃·▃·█·▅ beats 3/9	hard · avg 19%	▇ 1.5%	at its level · small run
SWE-bench Multilingual lab	45.8% ✎	▁··▁▃▃▅▇█▃▃▁ beats 10/60	hard · avg 18%	—	not counted yet: no other model in the same run to compare it with
LiveCodeBench v5 (Jul–Dec 2024) lab +1	82.1% ✎	█··········█ beats 1/1		—	not counted yet: no other model in the same run to compare it with

chat & writing▁▁▁▃▅▄▅█▆▆▇▃ #147 of 393 * · 6.4% of its score · 1 test

Beats INTELLECT-3 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.

test	result	vs the others in its run	difficulty	counts	why
LMArena (text)	1360.5 ✎	▁▁▁▂▃▃▃▆▆▇█▃ beats 204/379	avg 1268	▇▇▇ 6.4%	blunt

math▁▂▃▅▆▅▄█▄▁▁▁ #118 of 398 · 37% of its score · 7 tests

Beats Claude Sonnet 4.6, MiniMax-M3, MiMo-V2.5 (310B-A15B) here, though they rank higher overall. Trails Olmo 3.1 32B Think, Gemini 2.5 Flash, Qwen3-VL 32B Thinking here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
AIME 2025 lab +2	90.2% ✎	▁·▁▁▁▂▅▆█▇▅▁ beats 121/133	medium · avg 63%	▇▇▇▇ 9.2%	sharp · several runs
MathArena ArXivMath February 2026 (32 problems)	30.5% ✎	▁▁▁▁█▇▄▃▁▃·▃ beats 5/24	very hard · avg 9.1%	▇▇▇ 6.8%	at its level
MathArena ArXivMath January 2026 (23 problems)	48.9% ✎	▂··▄▂▄█▆▆▄▄█ beats 4/23	hard · avg 22%	▇▇ 6.0%	at its level
IMO-AnswerBench lab	77.2% ✎	▁▁·▁▁▁▄▅█▄▂▃ beats 17/42	medium · avg 42%	▇▇ 4.1%	typical
AIME 2026 lab	89.8% ✎	▁▁▁▂▁▂▆█▆▂·▁ beats 21/49	medium · avg 64%	▇▇ 3.8%	sharp
MathArena ArXivMath December 2025 (17 problems)	33.8% ✎	▃·▃▃█▁·▄·▁▁▁ beats 3/18	hard · avg 19%	▇ 3.7%	at its level
HMMT February 2025 lab +1	94.2% ✎	▁▁▁▃▁▅▅▃██▁▄ beats 24/31	easy · avg 72%	▇ 3.2%	small run

web searchnot ranked on this task yet · 2 tests

test	result	vs the others in its run	difficulty	counts	why
Humanity's Last Exam with tools lab	22.8% ✎	▂▂▃▁▄▂▃▆█▇▂▆ beats 5/54	hard · avg 16%	—	not counted yet: no other model in the same run to compare it with
BrowseComp lab	31.3% ✎	▂·▄▁▂▃▆▄█▂▃▄ beats 11/42	very hard · avg 10%	—	not counted yet: no other model in the same run to compare it with

other tests8 not in a task · 21% of its score

test	result	vs the others in its run	difficulty	counts	why
GPQA (subset unspecified) lab +1	80.5% ✎	▂▁▂▂▄▃▆▃▅█▄▃ beats 94/104	hard · avg 48%	▇▇ 6.2%	sharp · at its level
Humanity's Last Exam lab +1	18.5% ✎	▁▁▁▃▆▆▇██▆▁▃ beats 47/104	very hard · avg 5.0%	▇ 2.8%	small run
MMLU-Pro lab +1	83.8% ✎	▁▁·▁▁▁▃▂▄▅█▁ beats 161/188	medium · avg 58%	▇ 2.6%	small run
WideSearch corrected dataset	19.5% ✎	▃······▃▃▃█▃ beats 0/7	hard · avg 20%	▇ 2.4%	at its level · noisy
MultiChallenge lab +1	55.2% ✎	▁·▁·▁▃▆▅▃▃█▅ beats 18/29	hard · avg 27%	▇ 1.9%	at its level · small run
SkillsBench 87-task subset	20.3% ✎	▄··▄▄▄···█·█ beats 1/7	very hard · avg 8.2%	▇ 1.9%	at its level
MMLU-ProX lab +1	79.4% ✎	▁·▁▁▁▂▁▃▂█▅▄ beats 21/35	medium · avg 54%	▇ 1.7%	small run
IFBench prompt-level lab +1	73.4% ✎	█▄▄··▄·▄···▄ beats 2/6	medium · avg 55%	▇ 1.3%	small run

$ where to run it

Through an API

api · Amazon Bedrock · $0.28/M · $0.15 in / $0.65 out per 1M tokens · 262k context · docs.aws.amazon.com ↗ · price sourceStandard base token rate from models.dev. Provider-specific tiers, long-context surcharges, caching and promotions may differ.

llama.cpp · recommended

version

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 ↗
Run Nemotron 3 Super 120B-A12B · needs ~86.1 GB · from huggingface.co/lmstudio-community
chat in the terminal
llama cli -hf lmstudio-community/NVIDIA-Nemotron-3-Super-120B-A12B-GGUF:Q4_K_M
Or start a local server and open the address it prints (http://localhost:9931)
llama serve -hf lmstudio-community/NVIDIA-Nemotron-3-Super-120B-A12B-GGUF:Q4_K_M

LM Studio

version

Install LM Studio · Open the app once before using the lms command.
download it (macOS / Windows / Linux) ↗
Run Nemotron 3 Super 120B-A12B · needs ~86.1 GB · from huggingface.co/lmstudio-community
open in LM Studio ↗
Or download it in the terminal
lms get https://huggingface.co/lmstudio-community/NVIDIA-Nemotron-3-Super-120B-A12B-GGUF@Q4_K_M -y
Then run lms ls to see its name, and lms chat followed by that name

Ollama

Install Ollama
download it (macOS / Windows) ↗
Linux
curl -fsSL https://ollama.com/install.sh | sh
Run Nemotron 3 Super 120B-A12B · needs ~72.86 GB · from ollama.com
chat in the terminal
ollama run nemotron-3-super:120b

MLX (Mac)

version

Install MLX (Mac) · Apple silicon Macs only, text only.
with uv
uv tool install mlx-lm
or
pip install mlx-lm
Run Nemotron 3 Super 120B-A12B · needs ~68 GB, Mac only · from huggingface.co/mlx-community
chat in the terminal
mlx_lm.chat --model mlx-community/NVIDIA-Nemotron-3-Super-120B-A12B-4bit
Or start an OpenAI-compatible server at http://localhost:8080/v1
mlx_lm.server --model mlx-community/NVIDIA-Nemotron-3-Super-120B-A12B-4bit

Official weights and documentation: huggingface.co/nvidia ↗

$ how the score was worked out · NVIDIA vendor-reported benchm… 46%, NVIDIA — Nemotron-Cascade 2 t… 27%, MathArena ArXivMath / BrokenA… 17% and 2 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

source	results	share of the score	licence
NVIDIA vendor-reported benchmark scores for its own Nemotron models (Hugging Face model cards) · lab-reported	12	▇▇▇▇▇▇▇▇▇▇▇▇▇▇ 46%	Factual vendor-reported score (not licensed content)
NVIDIA — Nemotron-Cascade 2 technical report · lab-reported	8	▇▇▇▇▇▇▇▇ 27%	CC BY 4.0
MathArena ArXivMath / BrokenArXiv / ArXivLean model-output datasets	3	▇▇▇▇▇ 17%	CC BY-SA 4.0
LMArena leaderboard dataset	1	▇▇ 6%	CC BY 4.0
inclusionAI/Ling-3.0-flash model card	2	▇ 4%	MIT
prior · pull toward the middle		▇ 0%
weighted average		= 64.7

not counted (16)

nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16 model card	MMLU-Pro [correct]	83.7%	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-Super-120B-A12B-BF16 model card	AIME 2025 [correct]	90.2%	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-Super-120B-A12B-BF16 model card	HMMT February 2025 [correct]	93.7%	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-Super-120B-A12B-BF16 model card	GPQA (subset unspecified) [correct]	79.2%	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-Super-120B-A12B-BF16 model card	LiveCodeBench v5 (Jul–Dec 2024) [correct]	81.2%	fewer than three models have this test yet, or this source ran it for this model alone
nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16 model card	SciCode subproblems [correct]	42.0%	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-Super-120B-A12B-BF16 model card	Humanity's Last Exam [correct]	18.3%	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-Super-120B-A12B-BF16 model card	Terminal-Bench 2.0 [correct]	31.0%	fewer than three models have this test yet, or this source ran it for this model alone
nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16 model card	SWE-bench Multilingual [correct]	45.8%	fewer than three models have this test yet, or this source ran it for this model alone
nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16 model card	τ²-bench aggregate [correct]	61.2%	fewer than three models have this test yet, or this source ran it for this model alone
nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16 model card	IFBench prompt-level [correct]	72.6%	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 — Nemotron 3 Super technical report	BrowseComp [correct]	31.3%	fewer than three models have this test yet, or this source ran it for this model alone
NVIDIA — Nemotron 3 Super technical report	Humanity's Last Exam with tools [correct]	22.8%	fewer than three models have this test yet, or this source ran it for this model alone
NVIDIA — Nemotron 3 Super technical report	MMLU-ProX [correct]	79.4%	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 — Nemotron 3 Super technical report	MultiChallenge [correct]	55.2%	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 vendor-reported benchmark scores for its own Nemotron models (Hugging Face model cards)	LiveCodeBench v5 (Jul–Dec 2024) [correct]	82.1%	fewer than three models have this test yet, or this source ran it for this model alone

$ published speed measurements · 8

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)	Q4_K (ggml-org file Nemotron-3-Super-120B-Q4_K.gguf; llama-bench type label Q4_K)	19.94 tok/s	llama.cpp 04a65daab (8268), 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)	Q4_K (ggml-org file Nemotron-3-Super-120B-Q4_K.gguf; llama-bench type label Q4_K)	19.95 tok/s	llama.cpp 04a65daab (8268), 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)	Q4_K (ggml-org file Nemotron-3-Super-120B-Q4_K.gguf; llama-bench type label Q4_K)	19.83 tok/s	llama.cpp 04a65daab (8268), 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)	Q4_K (ggml-org file Nemotron-3-Super-120B-Q4_K.gguf; llama-bench type label Q4_K)	19.84 tok/s	llama.cpp 04a65daab (8268), 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)	Q4_K (ggml-org file Nemotron-3-Super-120B-Q4_K.gguf; llama-bench type label Q4_K)	19.45 tok/s	llama.cpp 04a65daab (8268), 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 ↗
AMD Ryzen AI Max+ 395 (Strix Halo), Radeon 8060S iGPU; Beelink GTR9 Pro 128GB	UD-IQ4_XS	18.24 tok/s	llama.cpp ac4cddeb0 (build_number 9592 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=999; threads=16; one sequence; file label r3 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 33 (2026-06-11, measured-local): Latest upstream Vulkan/RADV control for the 120B-class Nemotron capacity route; still direct-runnable on one 128GB Strix Halo system, but not a speed result.	source ↗
AMD Ryzen AI Max+ 395 (Strix Halo), Radeon 8060S iGPU; Beelink GTR9 Pro 128GB	UD-IQ4_XS	18.43 tok/s	llama.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 r3 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 36 (2026-06-05, measured-local): Latest/int-dot direct GGUF capacity route; 120B-class MoE runs on one 128GB Strix Halo system, but this is a capacity/current-model proof rather than a speed result.	source ↗
AMD Ryzen AI Max+ 395 (Strix Halo), Radeon 8060S iGPU; Beelink GTR9 Pro 128GB	UD-IQ4_XS	18.93 tok/s	llama.cpp 98d5e8ba8 (build_number 9544 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=999; threads=16; one sequence; file label r3 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 102 (2026-06-07, measured-local): Latest llama.cpp b9544 Vulkan/RADV 120B-class capacity control with explicit -dev Vulkan0; r3; latest build did not regress the Nemotron Super route.	source ↗

$ improve this page

Update model information Add a place to run it Add a test result Propose a benchmark Correct a benchmark score Edit 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 2026-10-11 02:20 UTC · sources and licences · json