NVIDIA launches Nemotron 3 for text and image moderation
NVIDIA introduces Nemotron 3 Content Safety, an open model that moderates text, images, and assistant responses in 12 languages. The company says it reaches 84% average accuracy in multimodal tests and reduces latency compared with larger safety models.
NVIDIA has introduced Nemotron 3 Content Safety, a 4-billion-parameter model that analyzes text, images, and assistant responses to detect unsafe content. Its goal is to address two problems that often appear together: moderation systems understand languages other than English less effectively, and they can miss the meaning when an image and text are interpreted separately.
The model is designed to monitor conversations, AI agents, documents, screenshots, and images. It can receive only text, only an image, or both at once, then issue a brief verdict on whether the content is safe.
Context changes meaning
A picture of a kitchen knife is not, by itself, a threat. If the accompanying text discusses preparing a meal, the combination is safe. But if it says the object will be used to hurt someone, the same knife becomes part of a possible threat.
That combined analysis also depends on language and culture. A religious symbol may have an acceptable meaning in one context and be associated with propaganda, hate, or discrimination in another. That is why translating each element separately is not always enough: the system has to understand what the combination means for the user and their environment.
How it works
Nemotron 3 Content Safety is based on Gemma-3 4B-IT, a vision-language model with a 128,000-token context window and support for more than 140 languages. NVIDIA fine-tuned it with a LoRA adapter, a technique that adds specific capabilities without completely modifying the base model.
The system can evaluate both the user's request and an assistant's generated response. This makes it possible to detect cases where the input appears safe but the response ends up including dangerous instructions or prohibited content.
In its low-latency mode, it returns labels such as:
- User request safety: safe
- Response safety: unsafe
- Categories: violence and crime planning or confession
The categories follow the taxonomy of the Aegis AI Content Safety Dataset v2, aligned with the ML Commons safety standard. You can also hide the categories and retain only the safety verdict.
Data in 12 languages
NVIDIA trained the model on human and synthetic data including text, real images, scanned documents, graphics, screenshots, and examples designed to test safety systems. The dataset covers risks such as:
- Harmful language, harassment, and self-harm
- Privacy violations
- Attempts to bypass a model's safeguards
- Crime planning
- Region-specific safety policies
The textual content was translated into 12 languages: English, Arabic, German, Spanish, French, Hindi, Japanese, Thai, Dutch, Italian, Korean, and Chinese. NVIDIA says synthetic data accounts for approximately 10% of the training, while most of the data comes from material created or reviewed by people.
Performance and availability
In tests cited by NVIDIA, the model reached an average accuracy of 84% in multimodal harmful-content evaluations and outperformed other open models of comparable size. It also maintained consistent results across the 12 languages used for training and showed zero-shot generalization, without specific training, in Portuguese, Swedish, Russian, Czech, Polish, and Bengali.
Speed is another goal. NVIDIA says it has approximately half the latency of larger multimodal safety models, including in average, median, and 99th-percentile measurements. The company says it can run on a GPU with 8 GB or more of VRAM, making it easier to use within an agent's workflow without delaying every action for too long.
The model is already available on Hugging Face for integration with Transformers or vLLM. It can work as a real-time filter, review large batches of documents and images, or control an agent's tool calls. NVIDIA also plans to offer it in April as NVIDIA NIM, a production-ready service that simplifies deployment and operation at scale.
For you, the most important change is that moderation is no longer limited to searching for specific words. A system like this attempts to assess who is saying what, in which language, alongside which image, and with what assistant response. Errors will remain, and the result will depend on the chosen data and policies, but this combination will become increasingly necessary as AI agents work with users, documents, and images from around the world.