AIPS BOOM

Nemotron-3-Content-Safety

3 min read

Nemotron 3 Content Safety

Type: NVIDIA model Tags: NVIDIA, Nemotron, content safety, multimodal safety, multilingual safety, NemoGuard, NeMo Guardrails, VLM safety, NIM, vLLM, Gemma, Build NVIDIA Related: Nemotron, NVIDIA-NemoGuard-NIMs, NeMo-Guardrails, NIM-for-Multimodal-Safety, NIM-for-Vision-Language-Models, NVIDIA-NIM, Nemotron-Content-Safety-Reasoning-4B-Experimental-NIM, Llama-3.1-Nemotron-Safety-Guard-8B-NIM, Llama-3.1-NemoGuard-8B-ContentSafety-NIM, NVIDIA-NemoGuard-JailbreakDetect-NIM, NVIDIA-AI-Enterprise, Triton-Inference-Server Sources: https://build.nvidia.com/nvidia/nemotron-3-content-safety/modelcard, https://build.nvidia.com/nvidia/nemotron-3-content-safety, https://huggingface.co/nvidia/Nemotron-3-Content-Safety, https://huggingface.co/blog/nvidia/nemotron-3-content-safety Last Updated: 2026-04-29

Summary

Nemotron 3 Content Safety is an NVIDIA multimodal, multilingual content-safety model for moderating prompts, optional images, and optional model responses. It extends NVIDIA’s text-oriented content-safety lineage into VLM and agent workflows by returning user-safety, response-safety, and optional safety-category judgments.

Detail

The Build NVIDIA model card describes Nemotron 3 Content Safety as a small language model based on Google’s Gemma-3-4B-it and fine-tuned by NVIDIA on multimodal and multilingual safety datasets. It is designed to moderate both inputs and outputs for LLMs and VLMs, including prompt-only, prompt+image, and prompt+image+response cases.

The model uses the safety taxonomy from NVIDIA’s Nemotron Content Safety Dataset V2 lineage and can return coarse safe/unsafe judgments or category-rich outputs. NVIDIA positions it as the multimodal counterpart to earlier text-only content-safety models such as Llama-3.1-NemoGuard-8B-ContentSafety-NIM and Llama-3.1-Nemotron-Safety-Guard-8B-NIM.

Architecturally, the model card identifies Gemma-3-4B-it as the base model, a decoder-only transformer architecture, SigLIP vision encoding for square images, about 4B parameters, and LoRA fine-tuning with merged weights. NVIDIA’s Hugging Face model card lists model version V1.1 and describes support for 12 languages, while the Build card lists a 128K context length and named support for English, Spanish, Mandarin, German, French, Hindi, Japanese, Arabic, and Thai.

Inputs and Outputs

  • Inputs: text prompts, optional images, and optional assistant responses.
  • Image input: image URL or base64-encoded image URL.
  • Use cases: multimodal input moderation, response moderation, VLM safety checks, agent-loop safety, and batch review of text/image content.
  • Outputs: text labels containing User Safety, optional Response Safety, and optional Safety Categories.
  • Safety categories: include violence, sexual content, criminal planning, illegal weapons, controlled substances, self-harm, identity hate, privacy/PII, harassment, threats, malware, fraud/deception, political misinformation/conspiracy, and other policy categories.

Deployment and Evaluation

The Build NVIDIA model card lists vLLM as the runtime engine, Linux as the operating system, and Ada Lovelace, Blackwell, and Hopper hardware compatibility, with L40S test hardware. The Hugging Face model card also shows Transformers and vLLM usage paths and lists H100, A100, and RTX PRO 6000 Blackwell Server Edition test hardware.

NVIDIA reports training, testing, and evaluation on multilingual text and image datasets, including NVIDIA ThreatOps data, Nemotron Safety Guard lineage data, Nemotron VLM Dataset V2, and synthetic generation. Evaluation covers external multimodal and multilingual safety benchmarks such as RTVLM, VLGuard, MM-SafetyBench, XSTEST, FigStep, Aegis 2, WildGuard, PolyGuard, RTP-LX, XSafety, MultiJail, and Aya Redteaming, plus false-positive checks on safe multimodal benchmarks including MMMU, DocVQA, and AI2D.

Connections

Resources

Graph View

Backlinks