NVIDIA: Nemotron 3.5 Content Safety

NVIDIA Nemotron 3.5 Content Safety is a compact 4B-parameter multimodal guardrail model from NVIDIA, fine-tuned from Google Gemma-3-4B. It is designed specifically for content moderation rather than general conversation, and it sits in the text category on api14. Its job is to act as a safety layer around other models, checking what goes into them and what comes back out.

The model accepts text and image inputs and returns text output, so it can moderate both text-only pipelines and vision-language pipelines. It moderates inputs to and responses from LLMs and VLMs, which means you can place it on either side of a generation call. The context window is 131072 tokens, and the maximum output is 117964 tokens. Supported parameters include temperature, top_p, top_k, min_p, frequency_penalty, presence_penalty, repetition_penalty, logit_bias, stop, seed, max_tokens, reasoning, and include_reasoning.

Practical use cases follow directly from that design. You can screen user prompts before they reach a chat or vision model, classify model responses before they are shown to end users, and run the same model over both text and image content in a single moderation pass. Because it handles images as well as text, it fits pipelines where users upload screenshots, photos, or other visual material alongside written input. Teams building assistants, community platforms, or internal tools can use it as a consistent policy checkpoint rather than writing separate classifiers for each modality.

Through api14 you call it with the OpenAI-compatible API using the model id "nvidia/nemotron-3.5-content-safety". Point your existing OpenAI client at the api14 base URL, send a chat completion request with that model id, and include the content you want moderated. The request parameters listed above are supported, so you can tune sampling with temperature, top_p, top_k, or min_p, control repetition with frequency_penalty, presence_penalty, and repetition_penalty, constrain output with max_tokens and stop, and use seed for reproducibility. The reasoning and include_reasoning parameters are available if you want the model's reasoning surfaced in the response.

Keep in mind what this model is for. It is a 4B guardrail model, not a general-purpose assistant, so tasks like long-form writing, complex reasoning, or open-ended chat are better served by a general text model. Use Nemotron 3.5 Content Safety when the goal is classification and policy enforcement over text and image inputs, and pair it with a generation model when you need both moderation and content production in the same pipeline.