NVIDIA Nemotron-H-56B-Base-8K is a large language model (LLM) developed by NVIDIA that is designed as a completion model for a given piece of text. It uses a hybrid model architecture that consists primarily of Mamba-2 and MLP layers combined with just ten Attention layers. The model features a context length of 8K. The supported languages include: English, German, Spanish, French, Italian, Korean, Portuguese, Russian, Japanese, and Chinese. For more detailed information on the model architecture, training, and evaluation, please see the project page and the technical report.

For best performance on a given task, users are encouraged to customize the model using the NeMo Framework suite of customization tools including Parameter-Efficient Fine-Tuning (P-tuning, Adapters, LoRA, and more), and Model Alignment (SFT, SteerLM, RLHF, and more) using NeMo-Aligner.

This model is for research and development only.

This model is part of the Nemotron-H Collection. You can find the models in this family here:

License/Terms of Use

GOVERNING TERMS: Use of this model is governed by the NVIDIA Internal Scientific Research and Development Model License.

Model Developer: NVIDIA

Model Dates:

October 2024 - March 2025

Data Freshness:

September 2024

The pretraining data has a cutoff date of September 2024.

Use Case:

This model is intended for developers and researchers building LLMs.

Release Date:

4/14/2025

References

[2504.03624] Nemotron-H: A Family of Accurate and Efficient Hybrid Mamba-Transformer Models

Model Architecture

Architecture Type: Hybrid Mamba-Transformer
Network Architecture: Nemotron-H

This model has 56B model parameters.

Input

Input Type(s): Text
Input Format(s): String
Input Parameters: One-Dimensional (1D): Sequences
Other Properties Related to Input: Context length up to 8K. Supported languages include German, Spanish, French, Italian, Korean, Portuguese, Russian, Japanese, Chinese and English.

Output

Output Type(s): Text
Output Format: String
Output Parameters: One-Dimensional (1D): Sequences

Our AI models are designed and/or optimized to run on NVIDIA GPU-accelerated systems. By leveraging NVIDIA’s hardware (e.g. GPU cores) and software frameworks (e.g., CUDA libraries), the model achieves faster training and inference times compared to CPU-only solutions.

Software Integration

Runtime Engine(s): NeMo 24.12
Supported Hardware Microarchitecture Compatibility: NVIDIA H100-80GB, NVIDIA A100
Operating System(s): Linux

Model Version

v1.0

Prompt Format

As this is a base model, no explicit prompt format is recommended or required.

Example

python

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

# Load the tokenizer and model
tokenizer  = AutoTokenizer.from_pretrained("nvidia/Nemotron-H-56B-Base-8K", trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained("nvidia/Nemotron-H-56B-Base-8K", torch_dtype=torch.bfloat16, trust_remote_code=True, device_map="auto")

prompt = "When was NVIDIA founded?"

outputs = model.generate(**tokenizer(prompt, return_tensors="pt", add_special_tokens=False).to(model.device))
print(tokenizer.decode(outputs[0]))

Training, Testing, and Evaluation Datasets

Training & Testing Datasets:

The training corpus for Nemotron-H-56B-Base-8K Base consists of English and multilingual text (German, Spanish, French, Italian, Korean, Portuguese, Russian, Japanese, Chinese and English), as well as code. Our sources cover a variety of document types such as: webpages, dialogue, articles, and other written materials. This model was also improved using synthetic data from Qwen (Built with Qwen). The corpus spans domains including legal, math, science, finance, and more. We also include a small portion of question-answering, and alignment style data to improve model accuracies.

Data Collection for Training & Testing Datasets: Hybrid: Automated, Human, Synthetic

Data Labeling for Training & Testing Datasets: Hybrid: Automated, Human, Synthetic

Evaluation Datasets

We used the datasets listed in the next section to evaluate Nemotron-H-56B-Base-8K Base.

Data Collection for Evaluation Datasets: Hybrid: Human, Synthetic

Data Labeling for Evaluation Datasets: Hybrid: Human, Synthetic, Automatic

Commonsense Understanding Evaluations:

ARC Challenge 25-shot	Hellaswag 10-shot	Winogrande 5-shot	CommonsenseQA 7-shot
94.97	89.00	84.45	86.73

ARC (Ai2 reasoning challenge)-Challenge - The challenge set of questions from a benchmark that contains grade-school level, multiple-choice science questions to assess question answering ability of language models. Dataset
Hellaswag - Tests the ability of a language model to correctly finish the provided context from a choice of possible options. Dataset
Winogrande - Tests the ability to choose the right option for a given sentence which requires commonsense reasoning. Dataset
CommonsenseQA - A multiple-choice question answering dataset that requires different types of commonsense knowledge to predict the correct answers. Dataset

Coding Evaluations:

MBPP (sanitized) 3-shot	MBPP+ 0-shot	HumanEval 0-shot	HumanEval+ 0-shot
77.82	67.20	60.37	54.27

MBPP (Mostly Basic Python Programming Problems) - Evaluates ability to generate solutions for Python programming tasks. Dataset
MBPP+ - Extended version of MBPP with additional validation. Dataset
HumanEval - Tests code generation and completion abilities in Python. Dataset

Math Evaluations:

GSM8K 8-shot CoT	MATH 4-shot CoT	MATH-Lvl 5 4-shot CoT	MATH-500 4-shot CoT
93.71	59.42	35.19	57.37

GSM8K (Grade School Math 8K) - Evaluates grade school level mathematical word problem solving. Dataset
MATH - Tests mathematical ability across multiple difficulty levels and various subjects including: Prealgebra, Algebra, Number Theory, Counting and Probability, Geometry, Intermediate Algebra, and Precalculus. Dataset
MATH Lvl 5 - Only the most difficult questions from the MATH dataset. Dataset
MATH-500 - Tests advanced mathematical problem solving across algebra, geometry, and calculus. Dataset

General Evaluations:

MMLU-Pro 5-shot CoT	MMLU 5-shot
60.51	84.21

MMLU Pro - Evaluates language understanding models across a broad range of challenging, reasoning-focused questions across 14 diverse domains. Dataset
MMLU - Tests knowledge across 57 subjects including science, humanities, math and more. Dataset

Potential Known Risks for Usage

The model was trained on data that contains toxic language, unsafe content, and societal biases originally crawled from the internet. Therefore, the model may amplify those biases and return toxic responses especially when prompted with toxic prompts. The model may generate answers that may be inaccurate, omit key information, or include irrelevant or redundant text producing socially unacceptable or undesirable text, even if the prompt itself does not include anything explicitly offensive.

The model demonstrates weakness to indirect prompt injection via some encodings, including Base16, Hex/ASCII, and Braille, though is more resilient than other similar models to injections using the more common Base64 vector.

Inference

Engine: NeMo
Test Hardware NVIDIA H100-80GB

Ethical Considerations

NVIDIA believes Trustworthy AI is a shared responsibility and we have established policies and practices to enable development for a wide array of AI applications. When downloaded or used in accordance with our terms of service, developers should work with their internal model team to ensure this model meets requirements for the relevant industry and use case and addresses unforeseen product misuse.

For more detailed information on ethical considerations for this model, please see the Responsible Use Guide available at http://nvidia.com/nemotron-responsible-use.

Please report security vulnerabilities or NVIDIA AI Concerns here.