The pre-training data has a cutoff date of June 25, 2025.
Description
NVIDIA-Nemotron-3-Nano-30B-A3B-Base-BF16 is a base large language model (LLM) trained from scratch by NVIDIA, with the next token prediction loss. It provides a good starting point for instruction fine-tuning.
This model is ready for commercial use.
What is Nemotron?
NVIDIA Nemotron™ is a family of open models with open weights, training data, and recipes, delivering leading efficiency and accuracy for building specialized AI agents.
We evaluated our model on the following benchmarks:
Task
Qwen3 30B-A3B-Base
NVIDIA-Nemotron-3-Nano-30B-A3B-Base-BF16
General Knowledge
MMLU (5-shot, acc)
81.07
78.56
MMLU-Pro (5-shot, CoT EM)
61.71
65.05
AGIEval-En (3/5-shot, CoT acc)
63.12
68.32
Code
HumanEval (0-shot)
70.73
78.05
MBPP-Sanitized (3-shot)
73.15
75.49
Math
GSM8K (8-shot, acc)
89.01
92.34
MATH (4-shot, acc)
61.14
82.88
MATH-500 (4-shot, avg@32)
55.08
78.63
Commonsense Understanding
ARC-Challenge (25-shot, acc_norm)
94.45
91.89
HellaSwag (10-shot, acc_norm)
83.14
85.56
OpenBookQA (0-shot, acc_norm)
44.80
46.20
PIQA (0-shot, acc_norm)
81.01
84.33
WinoGrande (5-shot, acc)
78.22
79.64
Reading Comprehension
RACE (0-shot, acc)
90.05
88.04
Multilingual
MMLU Global Lite (5-shot, avg acc)
76.84
74.47
MGSM (8-shot, avg acc)
82.53
83.00
Long Context
RULER (64K, 0-shot, acc)
63.55
87.50
RULER (128K, 0-shot, acc)
60.69
82.92
RULER (256K, 0-shot, acc)
Not Supported
75.44
RULER (512K, 0-shot, acc)
Not Supported
70.56
All evaluation results were collected via Nemo Evaluator SDK and LM Evaluation Harness. The open source container on LM Evaluation Harness packaged via NVIDIA's Nemo Evaluator SDK used for evaluations can be found here. A reproducibility tutorial along with all configs can be found in Nemo Evaluator SDK examples.
Deployment Geography: Global
Use Case
This model is intended for developers and researchers building instruction-following LLMs.
Our AI models are designed and 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.
The integration of foundation and fine-tuned models into AI systems requires additional testing using use-case-specific data to ensure safe and effective deployment. Following the V-model methodology, iterative testing and validation at both unit and system levels are essential to mitigate risks, meet technical and functional requirements, and ensure compliance with safety and ethical standards before deployment.
Use it with Transformers
The snippet below shows how to use this model with Huggingface Transformers (tested on version 4.57.3). We recommend using NeMo Framework 25.11.01 to ensure all required libraries are available.
Please note that the model supports up to a 1M context size, although the default context size in the Hugging Face configuration is 256k due to higher VRAM requirements.
from transformers import AutoTokenizer, AutoModelForCausalLM
model_name = "nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-Base-BF16"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
model_name,
torch_dtype=torch.bfloat16,
trust_remote_code=True,
device_map="auto"
)
prompt = "The capital of France is"
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(
**inputs,
max_new_tokens=32,
do_sample=False,
eos_token_id=tokenizer.eos_token_id
)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
Model Version(s)
v1.0
Training, Testing, and Evaluation Datasets
Data Modality: Text The total size: 10,648,823,153,919 Tokens Total number of datasets: 141 Dataset partition:Training [100%], testing [0%], validation [0%] Time period for training data collection: 2013 to May 1, 2025 Time period for testing data collection: 2013 to May 1, 2025 Time period for validation data collection: 2013 to May 1, 2025 Data Collection Method by dataset: Hybrid: Automated, Human, Synthetic
NVIDIA-Nemotron-3-Nano-30B-A3B-Base-BF16 is pre-trained on a large corpus of high-quality curated and synthetically-generated data. It is trained in the English language, as well as 19 other languages and 43 programming languages. Our sources cover a variety of document types such as: webpages, dialogue, articles, and other written materials. 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 accuracy. The model was trained for approximately 25 trillion tokens.
Alongside the model, we release our final pre-training data, as outlined in this section. For ease of analysis, there is a sample set that is ungated. For all remaining code, math and multilingual data, gating and approval is required, and the dataset is permissively licensed for model training purposes.
More details on the datasets and synthetic data generation methods can be found in the technical report NVIDIA Nemotron 3 Nano.
The English Common Crawl data was downloaded from the Common Crawl Foundation (see their FAQ for details on their crawling) and includes the snapshots CC-MAIN-2013-20 through CC-MAIN-2025-13. The data was subsequently deduplicated and filtered in various ways described in the Nemotron-CC paper. Additionally, we extracted data for fifteen languages from the following three Common Crawl snapshots: CC-MAIN-2024-51, CC-MAIN-2025-08, CC-MAIN-2025-18. The fifteen languages included were Arabic, Chinese, Danish, Dutch, French, German, Italian, Japanese, Korean, Polish, Portuguese, Russian, Spanish, Swedish, and Thai. As we did not have reliable multilingual model-based quality classifiers available, we applied just heuristic filtering instead—similar to what we did for lower quality English data in the Nemotron-CC pipeline, but selectively removing some filters for some languages that did not work well. Deduplication was done in the same way as for Nemotron-CC.
The GitHub Crawl was collected using the GitHub REST API and the Amazon S3 API. Each crawl was operated in accordance with the rate limits set by its respective source, either GitHub or S3. We collect raw source code and subsequently remove any having a license which does not exist in our permissive-license set (for additional details, refer to the technical report).
We use a considerable amount of synthetic data. Out of 10.6 trillion tokens, 3,534,013,958,278 tokens are synthetically generated.
We extracted data for fifteen languages from the following three Common Crawl snapshots: CC-MAIN-2024-51, CC-MAIN-2025-08, CC-MAIN-2025-18. The fifteen languages included were Arabic, Chinese, Danish, Dutch, French, German, Italian, Japanese, Korean, Polish, Portuguese, Russian, Spanish, Swedish, and Thai. As we did not have reliable multilingual model-based quality classifiers available, we applied just heuristic filtering instead—similar to what we did for lower quality English data in the Nemotron-CC pipeline, but selectively removing some filters for some languages that did not work well. Deduplication was done in the same way as for Nemotron-CC. Additionally, we used data from Wikipedia and FineWeb-2 (Penedo et al., 2025) for these fifteen languages.
Language
Total Tokens
Arabic
118,056,362,726
Danish
117,747,321,618
German
146,613,691,781
Spanish
469,156,575,409
French
139,982,002,289
Italian
298,858,370,174
Japanese
682,755,693,336
Korean
127,099,747,538
Dutch
89,041,592,681
Polish
105,356,493,147
Portuguese
243,249,275,089
Russian
185,314,014,057
Swedish
74,954,953,299
Thai
160,778,944,467
Chinese
211,007,236,689
We collect a total of 922,476,782,017 tokens of code in 43 different languages.
Language
Tokens
Assembly
750,628,764
C
42,657,300,868
C#
56,153,329,307
C++
67,773,701,658
CommonLisp
263,234,672
CSS
38,848,760,035
Cuda
400,222,993
Dart
3,816,960,470
Dockerfile
474,958,084
Fortran
1,105,049,387
Go
8,332,419,480
Haskell
1,294,613,669
HTML
69,082,117,487
Java
131,440,465,822
JavaScript
75,573,420,861
JSON
15,366,881,241
Julia
621,046,949
JupyterNotebook
2,241,893,197
Lua
4,146,420,802
Makefile
12,640,010,879
Markdown
64,796,743,311
Mathematica
320,504,225
OmniversePython
26,946,093
Pascal
1,625,013,876
Perl
1,575,314,434
PHP
61,575,339,005
Python
126,916,727,384
R
19,811,381,935
reStructuredText
1,779,876,391
Ruby
6,446,962,615
Rust
4,438,640,533
Scala
3,343,959,154
Shell
18,758,779,250
SQL
23,205,633,085
Swift
5,976,714,881
SystemVerilog
233,056,185
TeX
7,347,157,527
TypeScript
15,657,838,582
Verilog
811,884,369
VHDL
648,401,444
VisualBasic.NET
1,005,680,881
XML
12,616,779,741
YAML
10,574,010,491
Evaluation Dataset
Data Collection Method by dataset: Hybrid: Human, Synthetic
Labeling Method by dataset: Hybrid: Automated, Human, Synthetic
Inference
Engines: HF, vLLM, TRT-LLM
Test Hardware: NVIDIA A100 80GB, 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 Trustworthy AI 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.
We advise against circumvention of any provided safety guardrails contained in the Model without a substantially similar guardrail appropriate for your use case. For more details: Safety & Security.
For more detailed information on ethical considerations for this model, please see the Model Card++ Bias, Explainability, and Privacy Subcards.
Please report security vulnerabilities or NVIDIA AI Concerns here.
Citation
@misc{nvidia_nemotron_nano_v3_2025,
title = {{Nemotron 3 Nano}: Open, Efficient Mixture-of-Experts Hybrid {Mamba}-{Transformer} Model for {Agentic} Reasoning},
author = {{NVIDIA}},
year = {2025},
url = {https://arxiv.org/abs/2512.20848},
note = {Technical report}
}
