Nemotron 3.5 Content Safety: Customizable Multimodal Safety for Global Enterprise AI

Key Takeaways

→ Nemotron 3.5 Content Safety is built on Gemma 3 4B IT with a LoRA safety adapter and 128K context

→ It scores prompts, images and assistant responses together in one inference call

→ Supports 12 trained languages and zero-shot generalization to roughly 140

→ Custom policy enforcement plus optional THINK mode reasoning traces for audit

→ Averages about 85 percent harmful-content accuracy across multimodal benchmarks

→ Ships with an open safety dataset and runs under the NVIDIA Open Model License

NVIDIA released Nemotron 3.5 Content Safety on June 4, 2026, the latest in a line of guard models that has grown from an English text classifier into a family spanning modalities, languages and inference modes. This version unifies multimodal input, multilingual reach, custom enterprise policy and auditable reasoning in a single call.

Unified Multimodal Evaluation

Earlier versions scored text and images separately. Nemotron 3.5 takes the user prompt, an optional image and an optional assistant response as one context window and issues one verdict, catching violations that only emerge from the interaction between request, image and reply.

Global Language Coverage

The model keeps the 12-language explicit training coverage of its predecessors while inheriting strong zero-shot generalization to roughly 140 languages from the Gemma 3 base. That matters for deployments in markets where safety training data is sparse.

Custom Policy and THINK Mode

Enterprise deployments rarely share one safety taxonomy, so Nemotron 3.5 accepts a custom policy specification and reasons over it. An optional THINK mode emits a concise reasoning trace before the verdict, which regulators and reviewers can audit. The taxonomy follows Aegis 2.0: 13 core categories plus 10 subcategories.

Benchmarks and Latency

About 85 percent average harmful-content accuracy across evaluated multimodal benchmarks
96.5 percent on Multilingual Aegis across 12 languages
88.8 percent on RTP-LX, a combined 92.7 percent
3x lower end-to-end latency than an alternative multimodal safety model

Default mode latency is unchanged from Nemotron 3, so teams can keep real-time moderation on the fast path and run THINK-mode evaluation asynchronously.

Open Weights and an Open Dataset

The model is available on Hugging Face under the NVIDIA Open Model License, supports transformers, vLLM and SGLang, and ships with a multimodal, multilingual safety dataset that includes reasoning traces. It is also offered as a production-grade NIM microservice for teams that want a pre-packaged deployment.

Frequently Asked Questions

What is Nemotron 3.5 Content Safety?

It is NVIDIA's 4B-parameter multimodal, multilingual safety model released June 4, 2026, built on Gemma 3 4B IT and fine-tuned to classify prompts, images and assistant responses for policy violations.

Which languages does it support?

It is explicitly trained on 12 languages including English, French, Spanish, German, Chinese, Japanese, Korean and Arabic, and inherits zero-shot coverage of roughly 140 languages from its Gemma base model.

What is THINK mode?

THINK mode is an optional setting that makes the model output a short step-by-step reasoning trace before issuing its safe or unsafe verdict, which helps with compliance and human review.

Is Nemotron 3.5 Content Safety open?

Yes, the model is on Hugging Face under the NVIDIA Open Model License for research and commercial use, and NVIDIA also released the accompanying safety dataset.