Nemotron 3.5: Multimodal AI Security for Businesses

Nemotron 3.5: Content Security

Nemotron 3.5 represents a significant advancement in content security for artificial intelligences designed for enterprises. This version enhances the multimodal integration introduced by Nemotron 3, allowing for a unified evaluation of user prompts, images, and assistant responses. By processing these elements within a single context window, the model can now detect policy violations that arise from the interaction between text and image, or between request and response, in a single pass. This approach fills an important gap in multimodal security scenarios.

Global Language Coverage

Nemotron 3.5 maintains language coverage for 12 major languages, such as English, French, and Chinese, while inheriting zero-shot generalization capabilities in approximately 140 languages thanks to the Gemma 3 base model. This means that even in markets where training data is scarce, such as Southeast Asian languages or under-resourced African languages, deployments can benefit from multilingual transfer without requiring separate fine-tuning.

Custom Policy Application

One of the most significant architectural improvements in Nemotron 3.5 is the ability to apply custom security policies. Production deployments, whether in a healthcare platform or a financial services chatbot, can now integrate custom policy specifications. The model reasons about these policies when producing its verdict, rather than relying solely on the built-in taxonomy. This extends the work first introduced in Nemotron Content Safety Reasoning 4B to the multimodal and multilingual framework.

Reasoning Traces (THINK Mode)

Nemotron 3.5 offers an optional reasoning mode that allows each security verdict to be accompanied by an auditable reasoning trace. When activated, the model produces its reasoning step-by-step before delivering a final security label. This feature is particularly useful for compliance and auditing, as it provides a documented justification of the decision-making process.

Security Dataset

With the release of Nemotron 3.5, a security dataset is being published, marking a significant milestone in the field. Unlike most open-source security models, which typically do not provide their training or evaluation sets, this dataset is multimodal, multilingual, and includes security reasoning traces used to train the model.

Model Architecture

The content security model of Nemotron 3.5 is built on Google Gemma 3 4B IT, which consists of 4 billion parameters, offering a context window of 128K and strong vision-language reasoning. NVIDIA fine-tunes this base with a LoRA adapter, which installs targeted security classification behavior while keeping the model compact enough for real-time deployment on GPUs with 8 GB+ of VRAM.

Inference Interface

The inference interface of Nemotron 3.5 supports three output modes. Mode 1 provides a low-latency binary verdict, while Mode 2 adds security categories. Mode 3, or THINK Mode, includes a detailed reasoning trace. The security taxonomy follows the Aegis 2.0 framework, comprising 13 main categories and 10 detailed subcategories, aligned with the MLCommons security taxonomy.

Reasoning: An Accelerator for Security Classification

Reasoning plays a crucial role in accelerating content security classification, providing the context, customization, and accountability necessary for AI systems in production. It enables the model to dynamically interpret and apply domain-specific policies defined in natural language at the time of inference.

Optimizing Reasoning Traces

Although reasoning can introduce latency, Nemotron 3.5 optimizes this process by condensing reasoning chains into concise summaries, thereby limiting output tokens to enhance efficiency. This optimization occurs in two stages, similar to those implemented in the predecessor model Nemotron-Content-Safety-Reasoning-4B.

Data Sources

The dataset that powers Nemotron 3.5 is an evolution of the multimodal and multilingual mixes used for Nemotron 3, with additions targeting reasoning capabilities and custom policy application. Data sources include multilingual textual security data from the Nemotron Safety Guard Dataset v3, human-annotated multimodal data collected in English by NVIDIA and translated into 12 languages, as well as reasoning traces derived from the outputs of thought chains produced by larger teacher models.