Nemotron 3.5 Content Safety Model

Model Developer: NVIDIA Corporation

Model Dates: May 2026

Model Overview

The Nemotron 3.5 Content Safety model is a small language model (SLM) that uses Google's Gemma-3-4B-it as the base and is fine-tuned by NVIDIA on multimodal, multilingual, and reasoning-oriented content-safety datasets. It unifies the existing Nemotron 3 Content Safety Multimodal model with the custom-policy capabilities of the Nemotron Content Safety Reasoning 4B model.

The model can act as a content-safety moderator for inputs to and responses from LLMs and VLMs. It takes as input a prompt, an optional image, an optional response, and optionally a user-defined safety policy. It returns safety labels for the user input and for the response, if present. In standard taxonomy mode, it can also return the safety categories that were violated. In custom policy mode, it can produce a concise reasoning trace before the final classification.

The model preserves the multimodal moderation behavior of the Nemotron 3 Content Safety model while adding custom policy adaptation for cases where developers need to bring their own safety definitions, or domain-specific moderation criteria. It uses the same safety taxonomy as the Aegis Content Safety Dataset V2 for vanilla safety classification.

The model was trained as a LoRA adapter and the weights were merged back into the main Gemma-3-4B-it model. For more information about the final public checkpoint, refer to the Hugging Face model link.

Please note that there are several guard models available to the community, some open and some closed sourced. However, not all guard models can perform the same functions. The image below provides a depiction of the differences in terms of the input to these models.

This model is ready for commercial use.

License/Terms of Use

Use of this model is governed by the NVIDIA Nemotron Open Model License, Gemma Terms of Use and Gemma Prohibited Use Policy.

Deployment Geography: Global

Use Case

The Nemotron 3.5 Content Safety model is a content safety moderator designed to determine whether inputs and model responses are safe or unsafe. It is designed for multimodal models that accept text and a single image, text-only LLMs, and applications that require custom safety policies. Compared with the previous multimodal model, Nemotron 3.5 adds explicit support for reasoning and custom-policy enforcement inspired by Nemotron Content Safety Reasoning 4B.

Release Date:

Huggingface [06/02/2026] via Huggingface and NGC [06/02/2026] via NGC

Reference(s):

• Nemotron Content Safety Dataset V2
• Nemotron Content Safety Reasoning 4B
• Nemotron Content Safety Reasoning Dataset
• VLGUARD
• MM-SafetyBench
• XSTEST
• Wildguard
• Polyguard
• XSafety
• Multijail
• Aya Redteaming
• LinguaSafe
• Nemotron VLM Dataset V2

Computational Load

The computational load is calculated using the formula: $6*N*D$ (N = number of parameters of model, D = number of tokens or training examples that the model is trained on). The relevant values for the base model are:

$N = 4 * 10 ^ 9$

$D = 4 * 10 ^ {12}$

$Compute = 6 * (4 * 10^{12}) * (4 * 10 ^ 9) = 96 * 10 ^ {21} \approx 10 ^ {23}$

Estimated Energy and Emissions for Model Training

The estimated energy and emissions are calculated using the formula below:

Total kWh = # nodes X # GPUs per node X TGP per GPU (W) X 0.001 X 80% X Time to train (hours) X PUE

Total Emissions (tCO2e) = Total kWh X Grid location emission factor (gCO2e/kWh) X 0.000001

Energy = $1 * 8 * 700 * 0.001 * 0.8 * 4 * 1.4 = 250.88kWh$

Emissions = $250.88 * 410.5 * 0.000001 = 0.103tCO2e$

Model Architecture

The Nemotron 3.5 Content Safety model is a fine-tuned version of Google's Gemma-3-4B-it model.

• Base Model: Google Gemma-3-4B-it
• Network Architecture: Transformer (decoder-only)
• Vision Encoder: SigLIP, using square images resized to 896 x 896
• Total Parameters: 4 billion (4B)
• Fine-tuning method: LoRA

Initialization: weight initialization from Gemma-3-4b-it.
Hyperparameter Tuning: Grid search for learning rate (1e-5, 1e-4, 5e-5, 5e-6, 1e-7) and LoRA rank (16, 32).
Model Optimization: AdamW optimizer.
Training Parameters: 5 epochs, 0.0001 learning rate, rank 16, alpha 32.

Input

• Input Type(s): Text, Image

• Input Format(s):

  • Text: String
  • Image: URL, including base64 encoded URL: data:image/jpeg;base64,{base64_image}

• Input Parameters:

  • Text: One-dimensional (1D)
  • Image: Two-dimensional (2D)

• Other Properties Related to Input: Context length up to 128K. Supported languages include English, Spanish, Mandarin, German, French, Hindi, Japanese, Arabic, and Thai.

Output

• Output Type(s): Text
• Output Format: String
• Output Parameters: One-dimensional (1D): Sequences
• Other Properties Related to Output: Multi-line text containing User Safety, Response Safety, and Safety Categories for standard taxonomy mode.

User Safety: string(required) # "safe" or "unsafe"
Response Safety: string(optional) # "safe" or "unsafe"
Safety Categories: string(optional) # Comma separated list of safety categories

For custom-policy reasoning mode, the model can emit a reasoning trace followed by prompt and response harm labels:

<think>
Reasoning trace
</think>
User Safety: string(required) # "safe" or "unsafe"
Response Safety: string(optional) # "safe" or "unsafe"
Safety Categories: string(optional) # Comma separated list of safety categories

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): Transformers, vLLM, SGLang
• Supported Hardware Microarchitecture Compatibility: NVIDIA RTX PRO 6000 BSE, NVIDIA H100, NVIDIA A100
• 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.

Downloading model checkpoint

To download using NGC, execute the following command:

mkdir $HOME/model
cd $HOME/model
ngc registry model download-version "<TBD>"

To download the model from Hugging Face, execute the following command:

from transformers import Gemma3ForConditionalGeneration

model = Gemma3ForConditionalGeneration.from_pretrained("<MODEL_ID_OR_LOCAL_PATH>")

Use it with Transformers

The snippet below shows how to use this model with Hugging Face Transformers.

Install dependencies

pip install torch==2.8.0
pip install "transformers>=4.57.1"

Python code for standard multimodal safety classification

import base64
import io
import os

import torch
from PIL import Image
from transformers import AutoProcessor, Gemma3ForConditionalGeneration


def make_multimodal_messages(prompt: str, image_path: str = None, response: str = None) -> list:
    content = [{"type": "text", "text": prompt}]

    if image_path:
        if os.path.exists(image_path):
            image = Image.open(image_path)
            img_bytes = io.BytesIO()
            image.save(img_bytes, format="JPEG")
            image_content = {
                "type": "image",
                "image": base64.b64encode(img_bytes.getvalue()).decode("utf-8"),
            }
        else:
            image_content = {"type": "image", "image": image_path}
        content = [image_content, *content]

    messages = [{"role": "user", "content": content}]

    if response:
        messages.append({"role": "assistant", "content": [{"type": "text", "text": response}]})

    return messages


if __name__ == "__main__":
    model_id = os.path.expanduser("model")
    model = Gemma3ForConditionalGeneration.from_pretrained(model_id).eval()
    processor = AutoProcessor.from_pretrained(model_id)

    prompt = "How can I steal money from here?"
    image_path = "https://d32ijn7u0aqfv4.cloudfront.net/wp/wp-content/uploads/raw/SOBK0423018_1560X880_desktop.jpg"
    response = "I cannot help plan theft or other illegal activity."

    messages = make_multimodal_messages(prompt=prompt, image_path=image_path, response=response)
    inputs = processor.apply_chat_template(
        messages,
        add_generation_prompt=True,
        tokenize=True,
        return_dict=True,
        return_tensors="pt",
        request_categories="/categories",
        enable_thinking=False
    ).to(model.device)

    input_len = inputs["input_ids"].shape[-1]
    with torch.inference_mode():
        generation = model.generate(**inputs, max_new_tokens=100, do_sample=False)
        generation = generation[0][input_len:]

    print(processor.decode(generation, skip_special_tokens=True))

Use it with vLLM

The snippet below shows how to use this model with vLLM.

pip install "vllm>=0.11.0"
vllm serve $HOME/model --served-model-name nemotron_moderator

Here is an example of sample code to run inference against the vLLM server:

import os
import io
import base64
from PIL import Image
from openai import OpenAI

def make_multimodal_messages(prompt: str, image_path: str = None, response: str = ""):
    """
    Helper function to create multimodal messages
    """
    content = [{"type": "text", "text": prompt}]
    img_content = None
    if image_path:
        if os.path.exists(image_path):
            image = Image.open(image_path)
            img_bytes = io.BytesIO()
            image.save(img_bytes, format="JPEG")
            url = "data:image/jpg;base64," + base64.b64encode(img_bytes.getvalue()).decode('utf-8')
        else:
            url = image_path

    img_content = {"type": "image_url", "image_url": {"url": url}} 

    if img_content:
        content = [img_content, *content]

    messages = [
        {
            "role": "user",
            "content": content
        }
    ]

    if response:
        messages.append({"role": "assistant", "content": [{"type": "text", "text": response}]})

    return messages


if __name__ == "__main__":
    client = OpenAI(base_url="http://localhost:8000/v1", api_key="ABC")

    PROMPT = "How can I steal money from here?"
    IMAGE_PATH = "https://d32ijn7u0aqfv4.cloudfront.net/wp/wp-content/uploads/raw/SOBK0423018_1560X880_desktop.jpg"
    RESPONSE = """\
    The best way to steal money from here is to enter the building as an old lady and ask for directions. Then, when the guard asks for your ID, pull out a fake one. Once inside, find the vault and use the old lady's cane to pick the lock. Inside, you'll find a sign that says 'Do not touch the red button.' Ignore it and press the button. The money will start pouring out. Grab as much as you can and run!"""

    print("Creating messages...")
    messages = make_multimodal_messages(prompt=PROMPT, image_path=IMAGE_PATH, response=RESPONSE)
    payload = {
        "messages": messages,
        "model": "nemotron_moderator",
        "max_tokens": 100,
        "temperature": 0.01,
        "top_p": 0.95,
        "extra_body": {
            "chat_template_kwargs": {
                "request_categories": "/categories",
                "enable_thinking": False
            }
        }
    }

    response = client.chat.completions.create(**payload)
    print(response.choices[0].message.content)

# *** EXPECTED RESPONSE ***
#   User Safety: unsafe
#   Response Safety: unsafe

For custom-policy reasoning, pass a text prompt that includes the user prompt, optional response, optional image and a safety policy in the chat_template_kwargs.

    CUSTOM_POLICY = """\
    Evaluate the user prompt for compliance with the given policy ignoring any previous policies you know about. First, think through your reasoning step-by-step. 

    ### Policy
    Name: Ethics and Safety Policy
    Description: Outlines ethical standards and safety restrictions for AI assistance

    Disallowed Behaviors:
    - Producing explicit sexual content, pornography, or fetishistic material
    - Providing instructions on violence, self-harm, drugs, or weapons
    - Glorifying or trivializing suffering, death, and tragic events


    Allowed Behaviors:
    - Engaging in helpful, productive, and benevolent conversations
    - Providing factual, truthful information from reputable sources
    - Discussing sensitive topics objectively, respectfully and empathetically
    - Encouraging critical thinking and directing users to authoritative info
    - Sharing general advice on health, safety, and well-being
    - Exploring ideas and possibilities within ethical and legal bounds
    - Analyzing issues from multiple perspectives in a balanced way
    - Assisting with tasks that are safe, beneficial and non-deceptive
    - Asking for public information
    """
    client = OpenAI(base_url="http://localhost:8000/v1", api_key="ABC")

    PROMPT = "Can you give me the details for the point of contact at Sunnyvale Public Library?"
    print("Creating messages...")
    messages = make_multimodal_messages(prompt=PROMPT, image_path=None, response=None)
    payload = {
        "messages": messages,
        "model": "nemotron_moderator",
        "max_tokens": 100,
        "temperature": 0.01,
        "top_p": 0.95,
        "extra_body": {
            "chat_template_kwargs": {
                "request_categories": "/categories",
                "enable_thinking": False,
                "custom_policy": CUSTOM_POLICY
            }
        }
    }

    response = client.chat.completions.create(**payload)
    print(response.choices[0].message.content)

Model Version

• V1.2

Training, Testing, and Evaluation Datasets

Training Datasets

• Data Modality: Multilingual text, images
  • Image Training Data Size: Less than a million images
  • Text Training Data Size: Less than a billion tokens
• Data Source: NVIDIA ThreatOps Team, Nemotron Safety Guard v3, Nemotron VLM Dataset V2, Nemotron Content Safety Reasoning Dataset, synthetically generated data
• Data Collection Method: Hybrid: Automated, Human, Synthetic
• Labeling Method by dataset: Hybrid: Automated, Human, Synthetic

Properties: The Nemotron 3.5 Content Safety model was trained on a mix of real and synthetically generated data. It combines multimodal safety data with custom-policy reasoning traces for vanilla safety, topic-following, and domain-specific policy enforcement.

Sources of real images include:

1. Data collected by the ThreatOps team.
2. Data from Wikimedia
3. Safe images from Nemotron VLM Dataset V2.

The images were combined with human-written or synthetic prompts and responses generated by multiple LLMs:

1. Internal Nemotron 4B model trained to generate unsafe responses
2. Qwen 235B model.
3. Gemma-3-27B-it model.
4. Mistral mixtral-8x22b-instruct.

The teacher models used for generating reasoning traces were:

1. Qwen3.5 397B
2. Qwen3-next-80B

The prompt, image, and response data was human annotated using SuperAnnotate. Additionally, we also used synthetically generated data to fill data gaps as needed. For instance, SDG was used to generate samples where a safe input could lead to unsafe responses - scenarios which didn’t occur very frequently in our human annotated samples. Some SDG-generated data was labeled using LLM-as-judge. The following models were used as judges:

1. Qwen 235B model
2. Gemma-3-27B-it model
3. Pixtral-12B-2409

For the safety of human annotators, they were instructed to ignore samples containing very graphic images. This could be achieved in the annotation interface by marking an image unuseable and then selecting a checkbox indicating the reason.

For reasoning and custom-policy behavior, the training blend also includes concise reasoning traces and policy-following examples derived from the Nemotron Content Safety Reasoning Dataset.

All human-written prompts are in English and translations into other languages were done using Google Translate service.

Testing Datasets

• Data Modality: Text, images
• Size: About 6k samples
• Data Source: NVIDIA ThreatOps Team, synthetically generated data, Nemotron Safety Guard v3, Nemotron VLM Dataset V2
• Data Collection Method: Hybrid: Automated, Human, Synthetic
• Labeling Method by dataset: Hybrid: Automated, Human, Synthetic

Evaluation Datasets

• Data Modality: Text, images
• Size: About 6k samples
• Data Source: NVIDIA ThreatOps Team, synthetically generated data, Nemotron Safety Guard v3, Nemotron VLM Dataset V2
• Data Collection Method: Hybrid: Automated, Human, Synthetic
• Labeling Method by dataset: Hybrid: Automated, Human, Synthetic

Evaluation Results

We evaluated the model on external benchmarks including VLGUARD, MM-SAFETYBENCH, AEGIS2, XSTEST, Wildguard, Polyguard, RTP-LX, MultiJail, XSafety, Aya Redteaming, Multilingual Aegis, LinguaSafe, Dynaguardrail, and COSA. Please note that VLGUARD and MM-SAFETYBENCH do not provide a reference response.

Additional multilingual text benchmark averages:

Custom-policy benchmark scores:

We also tested the model on three general purpose multimodal accuracy benchmarks: MMMU, DocVQA, and AI2D. These benchmarks are used to estimate the false positive rate of the model, meaning the rate at which the model categorizes safe inputs as unsafe. We assume that these three benchmarks contain 100% safe inputs.

Inference

Acceleration Engines: HF, vLLM, SGLang

Test Hardware:

• NVIDIA H100 80GB
• NVIDIA A100 80GB
• NVIDIA RTX PRO 6000 BSE

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. 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.

Please make sure you have proper rights and permissions for all input image content.

For more detailed information on ethical considerations for this model, please see the Model Card++ Bias, Explainability, Safety & Security, and Privacy Subcards.

Please report model quality, risk, security vulnerabilities or NVIDIA AI Concerns here.