NVIDIA's Nemotron 3.5 Content Safety unifies multimodal, multilingual, custom policy AI safety
Single model evaluates text, images, and assistant responses with auditable reasoning and flexible policies.
NVIDIA has unveiled Nemotron 3.5 Content Safety, its latest enterprise-grade AI safety model that builds on the March 2026 Nemotron 3 release. The 4B-parameter model is the first to unify multimodal evaluation—accepting a user prompt, optional image, and optional assistant response in a single context window—and deliver a coherent safety verdict across all three inputs. This design closes a critical gap: policy violations that only become apparent from the interaction between text and image, or between request and response, are now caught in one pass. The model also expands language coverage through explicit training on 12 languages (English, French, Spanish, German, Chinese, Japanese, Korean, Arabic, Hindi, Russian, Portuguese, and Italian) and inherits zero-shot generalization to approximately 140 more languages from the Gemma 3 base model, making it suitable for global deployments without additional fine-tuning for low-resource languages.
A standout new feature is custom policy enforcement, allowing enterprises to supply their own safety taxonomy alongside each input. The model reasons over that custom policy when generating its verdict, rather than relying solely on a built-in taxonomy—essential for domain-specific use cases like healthcare, finance, or children's education. Additionally, an optional THINK mode produces auditable step-by-step reasoning traces for each safe/unsafe verdict, listing violated categories and explaining the interaction. When latency is critical, THINK mode can be disabled for fast binary outputs. Finally, NVIDIA released the Nemotron 3.5 Content Safety Dataset—a rare open-source artifact for multimodal safety that includes reasoning traces and covers multiple languages and modalities, enabling further research and fine-tuning.
- Unified multimodal evaluation catches cross-modal policy violations in a single pass across text, images, and assistant responses.
- Covers 12 languages explicitly and zero-shot generalizes to 140+ languages via the Gemma 3 base model.
- Supports custom enterprise policy taxonomies with auditable THINK-mode reasoning traces for compliance.
- Includes a released safety dataset with multimodal, multilingual reasoning traces for training and evaluation.
Why It Matters
NVIDIA's Nemotron 3.5 gives enterprises customizable, auditable, multimodal AI safety, crucial for compliant deployment across global, diverse use cases.