Provider logo

Nvidia Nemotron 3 Super 120B

nvidia/nemotron-3-super-120b-a12b
Provider logo

Nvidia Nemotron 3 Super 120B

nvidia/nemotron-3-super-120b-a12b

Nvidia's Nemotron 3 Super 120B A12B model from the March 2026 Nemotron 3 release. It uses a hybrid Mamba-Transformer MoE architecture and targets agentic and coding workloads with a 262K context window here.

Added Mar 1, 2026

Model weights

Context Window

262.1K

Max Output

16.4K

Avg output tokens (7d)

1.3K tokens

51%

Input Price (Auto)

$0.085/1M

Output Price (Auto)

$0.40/1M

Cache Read (Auto)

$0.042/1M

Capabilities

Benchmarks

Performance metrics and benchmarks

Sourced from Artificial Analysis.

Intelligence Index

25.7

Better than 63% of models compared

Coding Index

37.7

Better than 49% of models compared

Agentic Index

8.8

Better than 30% of models compared

Reasoning

GPQA Diamond

Graduate-level scientific reasoning

80.0%

Better than 70% of models compared

HLE

Humanity's Last Exam

20.8%

Better than 76% of models compared

IFBench

Instruction-following benchmark

71.5%

Better than 88% of models compared

T²-Bench Telecom

Conversational AI agents in dual-control scenarios

67.8%

Better than 61% of models compared

AA-LCR

Long context reasoning evaluation

60.3%

Better than 63% of models compared

GDPval-AA

Economically valuable tasks

9.9%

CritPt

Research-level physics reasoning

3.1%

Coding

SciCode

Python programming for scientific computing

36.0%

Better than 55% of models compared

Terminal-Bench Hard

Agentic coding and terminal use

28.8%

Better than 71% of models compared

Knowledge

AA-Omniscience Accuracy

Proportion of correctly answered questions

24.3%

AA-Omniscience Hallucination Rate

Rate of incorrect answers among non-correct responses

87.0%

Last updated Aug 16, 2026

Artificial Analysis

Providers

Choose explicit providers for this model. Auto routing remains available as the default option.

Loading provider options…

Compare Nvidia Nemotron 3 Super 120B with similar models from the same provider or model family.

Nvidia Nemotron 3 Super 120B Thinking

nvidia/nemotron-3-super-120b-a12b:thinking

Nvidia Nemotron 3 Super 120B with reasoning content enabled. Returns separate thinking content when requested.

Nvidia Nemotron Super 49B

nvidia/Llama-3.3-Nemotron-Super-49B-v1

Llama-3.3-Nemotron-Super-49B-v1 is a model which offers a great tradeoff between model accuracy and efficiency. Efficiency (throughput) directly translates to savings. Using a novel Neural Architecture Search (NAS) approach, we greatly reduce the model's memory footprint, enabling larger workloads, as well as fitting the model on a single GPU at high workloads (H200). This NAS approach enables the selection of a desired point in the accuracy-efficiency tradeoff. For more information on the NAS approach, please refer to this paper.

Nvidia Nemotron 3.5 Lightning

nvidia/nemotron-3.5-lightning

Nvidia Nemotron 3.5 Lightning is an open-weight 30B-A3B hybrid Mamba-Transformer mixture-of-experts model for high-throughput agentic, coding, tool-use, instruction-following, and long-context workloads. Thinking is disabled on this variant. Included in the NanoGPT subscription.

Nvidia Nemotron 3.5 Lightning Thinking

nvidia/nemotron-3.5-lightning:thinking

Nvidia Nemotron 3.5 Lightning is an open-weight 30B-A3B hybrid Mamba-Transformer mixture-of-experts model for high-throughput agentic, coding, tool-use, instruction-following, and long-context workloads. This variant enables its reasoning trace. Included in the NanoGPT subscription.

Nvidia Nemotron 3 Ultra 550B

nvidia/nemotron-3-ultra-550b-a55b

Nvidia's Nemotron 3 Ultra 550B A55B model from the Nemotron 3 family. It uses a hybrid Mamba-Transformer MoE architecture. Provider-specific context limits vary, with the longest current route supporting up to 1M context.

Nvidia Nemotron 3 Ultra 550B Thinking

nvidia/nemotron-3-ultra-550b-a55b:thinking

Nvidia's Nemotron 3 Ultra 550B A55B model from the Nemotron 3 family. It uses a hybrid Mamba-Transformer MoE architecture. Provider-specific context limits vary, with the longest current route supporting up to 1M context. Thinking enabled.