whatmodel.sh / Nemotron 3.5 Super VL 120B A12B

Nemotron 3.5 Super VL 120B A12B

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

Nemotron 3.5 Super VL 120B A12B is an open-weights language model from NVIDIA, released October 2026. Its published size is 120B parameters (12B active per token) and a 1M-token context window. Licence: open-mdw-1.1.

Catalogue identity notes

Precision is retained in each measurement: lab rows BF16; Super VL Arena rows NVFP4. Quantized deployment is not a separate checkpoint ranking.

Model information sources

Release date ↗

68.4±1.2#97 of 729 LLMsfrom 9 benchmark results across 2 sources

Score within the LMArena (text) comparison set only.

$ how it compares

The LLMs ranked either side of Nemotron 3.5 Super VL 120B A12B, out of 882 with a score, then more from NVIDIA.

#modelscorewhereRAMreleased
104Qwen3.6 27B Alibaba68±1api $1.20/M +local~18GB2026-04
105Ring-2.6-1T inclusionAI (Ant Group)68±2comparison only2026-05
106Nemotron 3.5 Super VL 120B A12B NVIDIA68±1local~254GB2026-10
107Nemotron 3.5 Bolt 74B A9B NVIDIA68±2local~157GB2026-10
108A.X K2 (688B-A33B) SK Telecom68±1local~1454GB2026-07
more from NVIDIA
102Nemotron 3 Ultra 550B-A55B NVIDIA69±1api $1.35/M +local~89GB2026-06
168Nemotron Cascade 2 30B-A3B NVIDIA65±1local~20GB2026-03
179Nemotron 3 Super 120B-A12B NVIDIA64±1api $0.28/M +local~73GB2026-03

see all LLMs · find the ones that fit your machine

$ what it's good at

agent tasks▁▁▃▆█▅▅▅▅▄▂▁ #66 of 310 · 27% of its score · 6 tests

Beats Gemini 3.6 Flash, LongCat 2.0 (1.6T-A48B), GPT-5 Pro here, though they rank higher overall.

testresultvs the others in its rundifficultycountswhy
SWE-bench Verified lab71.1% ✎▁▁▁▁▁▁▂▅█▆▁▁ beats 84/134hard · avg 21%▇▇▇▇▇ 12%sharp · at its level · small run
Terminal-Bench 2.1 lab60.7% ✎▁·▁·▁▁▄▃▅█▅█ beats 25/61hard · avg 15%▇▇▇ 7.4%sharp · at its level · small run
BrowseComp lab78.0% ✎▂·▃▁▂▂▆▄█▃▂▄ beats 35/44hard · avg 29%▇▇ 4.7%at its level · small run
τ³-bench Banking lab18.9% ✎▂▂·▅█▂▃▃▃▆▂▅ beats 17/24very hard · avg 6.3%▇ 2.2%blunt · small run
SWE-bench Multilingual lab67.7% ✎▁··▁▃▁▄▇█▂▂▁ beats 31/62hard · avg 18%—not counted yet: no other model in the same run to compare it with
SWE-bench Pro (split unspecified) lab51.8% ✎▃·▁▁▁▂▂▃█▇▃▁ beats 29/56very hard · avg 4.0%—not counted yet: no other model in the same run to compare it with
coding▁▁▁▃▆▇█▇▅▄▃▁ #90 of 466 · 46% of its score · 6 tests

Beats Grok 4.20 (0309 reasoning), GPT-5.4 mini, GPT-5.1 here, though they rank higher overall. Trails Llama 3.1 Nemotron Ultra 253B v1, GLM-4.7, GPT-5.4 nano here, though they rank lower overall.

testresultvs the others in its rundifficultycountswhy
Arena WebDev (Code Arena)1403.8 ✎▁▁▃▃▆█▄▇▅▅▁▁ beats 47/126avg 978▇▇▇▇▇▇▇▇▇▇▇▇ 29%sharp
SWE-bench Verified lab71.1% ✎▁▁▁▁▁▁▂▅█▆▁▁ beats 84/134hard · avg 21%▇▇▇▇▇ 12%sharp · at its level · small run
SciCode subproblems lab47.1% ✎▃··▃▃▃·▃▃█·▅ beats 5/10hard · avg 21%▇▇ 4.2%at its level · small run
LiveCodeBench v6 (window unspecified) lab88.2% ✎▂▃▄▄▅██▅▄▇▄▆ beats 56/66medium · avg 38%—not counted yet: no other model in the same run to compare it with
SWE-bench Multilingual lab67.7% ✎▁··▁▃▁▄▇█▂▂▁ beats 31/62hard · avg 18%—not counted yet: no other model in the same run to compare it with
SWE-bench Pro (split unspecified) lab51.8% ✎▃·▁▁▁▂▂▃█▇▃▁ beats 29/56very hard · avg 4.0%—not counted yet: no other model in the same run to compare it with
chat & writing▁▁▁▃▆▄▆█▆▆▇▃ #97 of 398 * · 16% of its score · 1 test

Beats LongCat 2.0 (1.6T-A48B), Inkling Small (266B-A12B) here, though they rank higher overall. Trails Gemma 4 12B, Kimi K2 Instruct (0905), Gemma 4 26B-A4B here, though they rank lower overall.

testresultvs the others in its rundifficultycountswhy
LMArena (text)1414.2 ✎▁▁▁▂▃▃▃▆▆▇█▃ beats 261/382avg 1259▇▇▇▇▇▇ 16%blunt
vision & documents▁▁▂▂▁▃▄▂▆▇█▄ #79 of 154 * · 10% of its score · 1 test

Beats DeepSeek V4 Flash Vision (Exp) here, though they rank higher overall. Trails Qwen3.5 9B, Gemma 4 12B, GPT-4.1 here, though they rank lower overall.

testresultvs the others in its rundifficultycountswhy
Arena vision (style controlled)1204.0 ✎▁▁▂▂▂▂▃▄▄▅█▇ beats 74/152avg 1061▇▇▇▇ 10%blunt
web search█··········█ #1 of 2 * · 4.7% of its score · 1 test
testresultvs the others in its rundifficultycountswhy
BrowseComp lab78.0% ✎▂·▃▁▂▂▆▄█▃▂▄ beats 35/44hard · avg 29%▇▇ 4.7%at its level · small run
other tests4 not in a task · 13% of its score
testresultvs the others in its rundifficultycountswhy
GPQA Diamond lab89.5% ✎▁▁▁▁▂▄▅▅█▇█▄ beats 267/315hard · avg 46%▇▇▇▇▇ 13%sharp · small run
IFBench (aggregation unspecified) lab70.9% ✎▃▇▂▄▅▅▂▄▄▆█▇ beats 57/83medium · avg 43%—not counted yet: no other model in the same run to compare it with
MMLU-ProX lab76.9% ✎▁·▁▁▁▂▁▃▃█▅▄ beats 15/36medium · avg 58%—not counted yet: no other model in the same run to compare it with
MMLongBench-Doc lab62.8% ✎▃···▃···▃▅█▅ beats 9/9—not counted yet: no other model in the same run to compare it with

$ where to run it

Other ways to run it

  1. Run it as published · needs ~240 GB, NVIDIA only · from huggingface.co/nvidia (official)
    vllm serve nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-BF16 \ --served-model-name nemotron-3.5-super-vl \ --tensor-parallel-size 4 \ --trust-remote-code \ --gpu-memory-utilization=0.85 \ --allowed-local-media-path=/ \ --max-num-seqs=32 \ --limit-mm-per-prompt '{"video": 1, "image": 1}' \ --video-pruning-rate 0.5 \ --media-io-kwargs '{"video": {"fps": 2, "num_frames": 256}}' \ --enable-auto-tool-choice \ --tool-call-parser qwen3_coder \ --reasoning-parser nemotron_v3 \ --speculative_config.method mtp \ --speculative_config.num_speculative_tokens 3

Official weights and documentation: huggingface.co/nvidia ↗

$ how the score was worked out · LMArena leaderboard dataset 56%, NVIDIA vendor-reported benchm… 44%

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
LMArena leaderboard dataset3CC BY 4.0
NVIDIA vendor-reported benchmark scores for its own Nemotron models (Hugging Face model cards) · lab-reported6Factual vendor-reported score (not licensed content)
prior · pull toward the middle
weighted average= 68.4

not counted (6)

NVIDIA vendor-reported benchmark scores for its own Nemotron models (Hugging Face model cards)MMLU-ProX [correct]76.9%fewer than three models have this test yet, or this source ran it for this model alone
NVIDIA vendor-reported benchmark scores for its own Nemotron models (Hugging Face model cards)LiveCodeBench v6 (window unspecified) [correct]88.2%fewer than three models have this test yet, or this source ran it for this model alone
NVIDIA vendor-reported benchmark scores for its own Nemotron models (Hugging Face model cards)IFBench (aggregation unspecified) [correct]70.9%fewer than three models have this test yet, or this source ran it for this model alone
NVIDIA vendor-reported benchmark scores for its own Nemotron models (Hugging Face model cards)SWE-bench Multilingual [correct]67.7%fewer than three models have this test yet, or this source ran it for this model alone
NVIDIA vendor-reported benchmark scores for its own Nemotron models (Hugging Face model cards)SWE-bench Pro (split unspecified) [correct]51.8%fewer than three models have this test yet, or this source ran it for this model alone
NVIDIA vendor-reported benchmark scores for its own Nemotron models (Hugging Face model cards)MMLongBench-Doc [correct]62.8%fewer than three models have this test yet, or this source ran it for this model alone

$ 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