Model

Nemotron-H 56B

OPEN

NVIDIA · Nemotron-H family · released Apr 14, 2025

Hybrid Mamba-2/Transformer base model, up to 3x faster inference than similarly-sized pure Transformers; non-commercial research license only.

ReasoningCodingVisionFunction callingTool useAgentic

883.7

Elo · rank #240

Parameters

56B

Active params

Undisclosed

Context

8K tokens

Architecture

Hybrid Mamba-2 + Transformer (54 Mamba-2 layers, 54 MLP layers, 10 self-attention layers of 118 total)

License

NVIDIA Internal Scientific Research and Development Model License (non-commercial, internal R&D only)

Languages

10+

API price (in/out)

No hosted API

Modalities

text

Benchmark results

Bar shows position within the tracked field; marker = field best

ARC-ChallengeReasoning95.0%#7

best: Llama 3.1 405B · 96.9%

GSM8KMath93.7%#19

best: Llama 3.1 405B · 96.8%

HellaSwagReasoning89.0%#13

best: Claude 3 Opus · 95.4%

HumanEval+Coding54.3%#25

best: Mistral Small 3.2 (24B) · 92.9%

HumanEvalCoding60.4%#101

best: Claude Opus 4.5 · 99.4%

MATH-500Math59.4%#114

best: GPT-5 · 99.4%

MBPP+Coding67.2%#20

best: Llama 3.1 405B · 88.6%

MBPPCoding77.8%#20

best: Llama-3.3-Nemotron-Super-49B v1 (Reasoning On) · 91.3%

MMLU-ProKnowledge60.5%#114

best: Claude Fable 5 · 91.5%

MMLU (STEM)Knowledge80.6%#4

best: Falcon-H1 34B · 83.6%

MMLUKnowledge84.2%#64

best: OpenAI o3 · 92.9%

OpenBookQAReasoning48.6%#29

best: Claude 1 · 90.8%

PIQAReasoning85.0%#9

best: GPT-4o mini · 93.1%

WinoGrandeReasoning84.5%#12

best: PaLM 2 · 90.9%

Run it locally

VRAM @ Q4

—

VRAM @ FP16

—

Fits on (Q4)

Multi-node cluster required

Non-commercial research-only license; no consumer-GPU quantization confirmed.

Quantizations

—

Nemotron-H family

Elo progression across releases

Nemotron-H 56BApr 2025883.7

Resources

HFnvidia/Nemotron-H-56B-Base-8K GHNVIDIA-NeMo/Nemotron DOCSresearch.nvidia.com/labs/adlr/nemotronh

API price weights · each benchmark row carries its own source badge (see methodology)