NVIDIA Unveils Nemotron 3.5 Content Safety

NVIDIA has unveiled Nemotron 3.5 Content Safety, an evolution of the existing Nemotron 3. Built on Google's Gemma 3 4B IT, this model offers the following key features.

• Unified Multimodal Evaluation: Processes user prompts, images, and assistant responses within a single context, detecting violations arising from interactions between text and images in a single pass.
• Global Language Support: The model has been explicitly trained on 12 languages, including Korean, and, building on Gemma 3's performance, achieves strong zero-shot generalization across approximately 140 languages.
• Custom Policy Enforcement: When enterprises input their own uniquely defined safety policies into the model, the model can make judgments based on those policies, enabling tailored responses for specific industries (finance, healthcare, etc.).
• Reasoning Process Output (THINK Mode): Optionally supports THINK mode, which outputs the model's step-by-step reasoning process, enhancing the auditability of safety determination results.

NVIDIA has also released the multimodal/multilingual safety dataset and reasoning traces used to train the model, supporting safety research within the open-source ecosystem.