NVIDIA Nemotron 3: Efficient and Open Intelligence
Paper • 2512.20856 • Published • 43
Read our How to Run Nemotron 3 Nano Guide!
See Unsloth Dynamic 2.0 GGUFs for our quantization benchmarks.
- Note
<think>and</think>are separate tokens, so use--specialif needed. - You can also fine-tune the model with Unsloth.
NVIDIA-Nemotron-3-Nano-4B-BF16
Model Developer: NVIDIA Corporation
Model Dates:
Dec 2025 - Jan 2026
Data Freshness:
September 2024
The pretraining data has a cutoff date of September 2024.
Model Overview
NVIDIA-Nemotron-3-Nano-4B-BF16 is a small language model (SLM) trained from scratch by NVIDIA, and designed as a unified model for both reasoning and non-reasoning tasks. It responds to user queries and tasks by first generating a reasoning trace and then concluding with a final response. The model's reasoning capabilities can be controlled via a system prompt. If the user prefers the model to provide its final answer without intermediate reasoning traces, it can be configured to do so, albeit with a slight decrease in accuracy for harder prompts that require reasoning. Conversely, allowing the model to generate reasoning traces first generally results in higher-quality final solutions to queries and tasks.
The model has been compressed from NVIDIA-Nemotron-Nano-9B-v2 using the Nemotron Elastic framework. The details of the parent model NVIDIA-Nemotron-Nano-9B-v2 can be found in (Nemotron-H tech report). The model uses a hybrid architecture consisting primarily of Mamba-2 and MLP layers combined with just four Attention layers.
The supported languages include: English. Improved using Qwen.
This model is ready for commercial use.
License/Terms of Use
Governing Terms: Use of this model is governed by the NVIDIA Nemotron Open Model License.
Deployment Geography: Global
Use Case
NVIDIA-Nemotron-3-Nano-4B is an edge-ready small language model intended for Agentic AI in edge platforms (Jetson Thor, GeForce RTX, DGX Spark). It targets key-uses including AI gaming NPCs (teammates / companions), local voice assistants (for devices, apps, and games), and IoT automation. It is to be used in English and coding languages.
Release Date: 3/16/2026
Huggingface 3/16/2026 via https://huggingface.co/
References
- NVIDIA Nemotron Nano 2: An Accurate and Efficient Hybrid Mamba-Transformer Reasoning Model
- Nemotron Elastic: Towards Efficient Many-in-One Reasoning LLMs
- NVIDIA Nemotron 3: Efficient and Open Intelligence
- Nemotron 3 Nano: Open, Efficient Mixture-of-Experts Hybrid Mamba-Transformer Model for Agentic Reasoning
- Nemotron 3 Super: Open, Efficient Mixture-of-Experts Hybrid Mamba-Transformer Model for Agentic Reasoning
Model Architecture
- Architecture Type: Mamba2-Transformer Hybrid
- Network Architecture: Nemotron-Hybrid
- This model was compressed from nvidia/NVIDIA-Nemotron-Nano-9B-v2
- Number of model parameters 3.97 x 10^9
Input
- Input Type(s): Text
- Input Format(s): String
- Input Parameters: One-Dimensional (1D): Sequences
- Other Properties Related to Input: Context length up to 262K. Supported languages include English.
Output
- Output Type(s): Text
- Output Format: String
- Output Parameters: One-Dimensional (1D): Sequences
- Other properties Related to Output: Sequences up to 262K
Our 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.
Software Integration
- Runtime Engine(s): NeMo 25.07
- Supported Hardware Microarchitecture Compatibility: NVIDIA A10G, NVIDIA H100-80GB, NVIDIA A100, GeForce RTX
- Operating System(s): Linux
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 4.48.3).
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM
# Load tokenizer and model
tokenizer = AutoTokenizer.from_pretrained("nvidia/NVIDIA-Nemotron-3-Nano-4B")
model = AutoModelForCausalLM.from_pretrained(
"nvidia/NVIDIA-Nemotron-3-Nano-4B",
torch_dtype=torch.bfloat16,
trust_remote_code=True,
device_map="auto"
)
messages = [
{"role": "system", "content": <system_prompt>},
{"role": "user", "content": "Write a haiku about GPUs"},
]
tokenized_chat = tokenizer.apply_chat_template(
messages,
tokenize=True,
add_generation_prompt=True,
return_tensors="pt"
).to(model.device)
outputs = model.generate(
tokenized_chat,
max_new_tokens=32,
eos_token_id=tokenizer.eos_token_id
)
print(tokenizer.decode(outputs[0]))
temperature=1.0 and top_p=0.95 are recommended for reasoning tasks, while temperature=0.6 and top_p=0.95 are recommended for tool calling.
If you’d like to use reasoning off, add enable_thinking=False to apply_chat_template(). By default, enable_thinking is set to be True.
messages = [
{"role": "system", "content": <system_prompt>},
{"role": "user", "content": "Write a haiku about GPUs"},
]
tokenized_chat = tokenizer.apply_chat_template(
messages,
tokenize=True,
enable_thinking=False,
add_generation_prompt=True,
return_tensors="pt"
).to(model.device)
outputs = model.generate(
tokenized_chat,
max_new_tokens=32,
eos_token_id=tokenizer.eos_token_id
)
print(tokenizer.decode(outputs[0]))
Use it with vLLM
We need vllm>=0.15.1 for this model. If you are on Jetson Thor or DGX Spark, please use this vllm container.
pip install -U "vllm>=0.15.1"
Download the custom parser from the Hugging Face repository.
wget https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-4B-BF16/resolve/main/nano_v3_reasoning_parser.py
Launch a vLLM server using the custom parser.
vllm serve nvidia/NVIDIA-Nemotron-3-Nano-4B-BF16 \
--served-model-name nemotron3-nano-4B-BF16\
--max-num-seqs 8 \
--tensor-parallel-size 1 \
--max-model-len 262144 \
--port 8000 \
--trust-remote-code \
--mamba_ssm_cache_dtype float32 \
--enable-auto-tool-choice \
--tool-call-parser qwen3_coder \
--reasoning-parser-plugin nano_v3_reasoning_parser.py \
--reasoning-parser nano_v3
Access the hosted API using a python client.
from openai import OpenAI
import asyncio
from openai import AsyncOpenAI
# NOTE: Streaming is preferred for better performance and resource efficiency.
# It allows you to start processing responses as they arrive, reducing latency.
# Synchronous example (non-streaming)
client = OpenAI(
api_key="your-nvapikey",
base_url="base-url"
)
response = client.chat.completions.create(
model="nemotron3-nano-4B-BF16",
messages=[
{
"role": "user",
"content": "Hello!"
}
],
temperature=0.7,
max_tokens=256,
top_p=0.7,
stream=false
)
print(response.choices[0].message.content)
Use it with TRT-LLM
Launch the model using TRT-LLM
docker run -v /home/root/.cache/huggingface/:/root/.cache/huggingface/ --rm --ulimit memlock=-1 --ulimit stack=67108864 --gpus=all --ipc=host --network host -d -e MODEL=NVIDIA-Nemotron-3-Nano-4B-BF16 -e HF_TOKEN=$HF_TOKEN nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc6 bash -c '
cat > /tmp/extra-llm-api-config.yml <<EOF
kv_cache_config:
dtype: "auto"
enable_block_reuse: false
cuda_graph_config:
max_batch_size: 32
enable_padding: true
disable_overlap_scheduler: true
moe_config:
backend: CUTLASS
EOF
trtllm-serve \
NVIDIA-Nemotron-3-Nano-4B-BF16 \
--host 0.0.0.0 \
--port 8123 \
--max_batch_size 32 \
--extra_llm_api_options /tmp/extra-llm-api-config.yml '
Access the hosted endpoint using curl command.
curl http://localhost:8123/v1/chat/completions -H "Content-Type: application/json" -d '{
"model": "NVIDIA-Nemotron-3-Nano-4B-BF16",
"messages": [
{
"role": "user",
"content": "Where is New York?"
}
],
"max_tokens": 1024,
"top_p": 1.0
}' -w "\n"
Model Version
- v1.0
Training, Testing, and Evaluation Datasets
Training datasets
- Data Modality: Text
- Text Training Data Size: More than 10 Trillion Tokens
- Train/Test/Valid Split: We used 100% of the corpus for pre-training and relied on external benchmarks for testing.
- Data Collection Method by dataset: Hybrid: Automated, Human, Synthetic
- Labeling Method by dataset: Hybrid: Automated, Human, Synthetic
Properties: The post-training corpus for NVIDIA-Nemotron-3-Nano-4B consists of English and multilingual text (German, Spanish, French, Italian, Korean, Portuguese, Russian, Japanese, Chinese and English). Our sources cover a variety of document types such as: webpages, dialogue, articles, and other written materials. The corpus spans domains including code, legal, math, science, finance, and more. We also include a small portion of question-answering, and alignment style data to improve model accuracies. For several of the domains listed above we used synthetic data, specifically reasoning traces, from DeepSeek R1/R1-0528, Qwen3-235B-A22B, Nemotron 4 340B, Qwen2.5-32B-Instruct-AWQ, Qwen2.5-14B-Instruct, Qwen 2.5 72B.
More details on the datasets and synthetic data generation methods can be found in the technical report NVIDIA Nemotron Nano 2: An Accurate and Efficient Hybrid Mamba-Transformer Reasoning Model .
Public Datasets
Private Non-publicly Accessible Datasets of Third Parties
| Dataset |
|---|
| Global Regulation |
| Workbench |
Online Dataset Sources
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).
| Dataset | Modality | Dataset Size (Tokens) | Collection Period |
|---|---|---|---|
| English Common Crawl | Text | 3.360T | 4/8/2025 |
| Multilingual Common Crawl | Text | 812.7B | 5/1/2025 |
| GitHub Crawl | Text | 747.4B | 4/29/2025 |
| English Common Crawl 1.1 | Text | Not disclosed | 10/2/2025 |
| Dataset | Collection Period |
|---|---|
| Problems in Elementary Mathematics for Home Study | 4/23/2025 |
| GSM8K | 4/23/2025 |
Evaluation Dataset:
- Data Collection Method by dataset: Hybrid: Human, Synthetic
- Labeling Method by dataset: Hybrid: Automated, Human, Synthetic
Evaluation Results:
Benchmark Results (Reasoning On)
We evaluated our model in **Reasoning-On** mode across these benchmarks.
| Benchmark | NVIDIA-Nemotron-3-Nano-4B-BF16 |
|---|---|
| AIME25 | 78.5 |
| MATH500 | 95.4 |
| GPQA | 53.2 |
| LCB | 51.8 |
| BFCL v3 | 61.1 |
| IFEVAL-Prompt | 87.9 |
| IFEVAL-Instruction | 92 |
| Tau2-Airline | 33.3 |
| Tau2-Retail | 39.8 |
| Tau2-Telecom | 33 |
We also evaluated our model in **Reasoning-off** mode across these benchmarks
| Benchmark | NVIDIA-Nemotron-3-Nano-4B-BF16 |
|---|---|
| BFCL v3 | 61.1 |
| IFBench-Prompt | 43.2 |
| IFBench-Instruction | 44.2 |
| Orak | 22.9 |
| IFEval-Prompt | 82.8 |
| IFEval-Instruction | 88 |
| HaluEval | 62.2 |
| RULER (128k) | 91.1 |
| Tau2-Airline | 28.0 |
| Tau2-Retail | 34.8 |
| Tau2-Telecom | 24.9 |
| EQ-Bench3 | 63.2 |
All evaluations were done using NeMo-Skills & Orak. For Orak we evaluated on three games (Super Mario, Darkest Dungeon & StarDew Valley)
Inference
- Engines: HF, vLLM, llama-cpp, TRT-LLM, SGLang
- Test Hardware: NVIDIA GeForce RTX, H100 80GB, DGX Spark, Jetson Thor/Orin Nano
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 and Explainability Subcards.
For more detailed information on ethical considerations for this model, please see the Model Card++ Bias, and Privacy Subcards.
Please report security vulnerabilities or NVIDIA AI Concerns here.
- Downloads last month
- 29,578
Model tree for unsloth/NVIDIA-Nemotron-3-Nano-4B
Base model
nvidia/NVIDIA-Nemotron-Nano-12B-v2-Base Finetuned
nvidia/NVIDIA-Nemotron-Nano-12B-v2 Finetuned
nvidia/NVIDIA-Nemotron-Nano-9B-v2 Finetuned
nvidia/NVIDIA-Nemotron-3-Nano-4B-BF16Papers for unsloth/NVIDIA-Nemotron-3-Nano-4B
Nemotron 3 Nano: Open, Efficient Mixture-of-Experts Hybrid Mamba-Transformer Model for Agentic Reasoning
Paper • 2512.20848 • Published • 42
Nemotron Elastic: Towards Efficient Many-in-One Reasoning LLMs
Paper • 2511.16664 • Published • 29
Nemotron-H: A Family of Accurate and Efficient Hybrid Mamba-Transformer Models
Paper • 2504.03624 • Published • 17
Nemotron-CC: Transforming Common Crawl into a Refined Long-Horizon Pretraining Dataset
Paper • 2412.02595 • Published • 8