NVIDIA releases Nemotron 3.5 Content Safety, unifying multimodal input, multilingual support, custom policy enforcement, and reasoning traces in a single 4B-parameter model

Nemotron 3.5 Content Safety evaluates user prompts, optional images, and optional assistant responses together in a single context window to catch policy violations that emerge from interactions between text and image or between request and response, rather than scoring each independently.

The model maintains explicit training coverage across 12 languages (English, French, Spanish, German, Chinese, Japanese, Korean, Arabic, Hindi, Russian, Portuguese, and Italian) while inheriting zero-shot generalization across approximately 140 languages from the Gemma 3 base model, benefiting deployments in language-sparse markets.

Custom policy specifications can be supplied at inference time, allowing the model to enforce domain-specific safety rules—such as suppressing irrelevant categories or defining proprietary risk categories—rather than defaulting to a single universal safety taxonomy.

An optional reasoning mode outputs step-by-step justification before delivering a final safe/unsafe verdict, enabling compliance documentation and human review; when latency is the primary constraint, this mode can be disabled to return to low-latency binary verdicts. NVIDIA is releasing the Nemotron 3.5 Content Safety Dataset, which is multimodal, multilingual, and includes the reasoning traces used for model training.