NVIDIA: Nemotron 3.5 Content Safety (free)

NVIDIA Nemotron 3.5 Content Safety is a guardrail model from NVIDIA, released by the provider on 2026-06-04. It is a compact 4B-parameter multimodal model, fine-tuned from Google Gemma-3-4B. Its job is moderation: it evaluates both the inputs sent to LLMs and VLMs and the responses those models generate, so it sits alongside a generation model rather than replacing one. On api14 it is listed under the text category and carries the model id nvidia/nemotron-3.5-content-safety:free.

The model accepts text and image inputs and returns text output. That multimodal input support matters for moderation pipelines, because a safety check often needs to consider an attached image alongside the prompt text rather than the text alone. It has a context length of 128,000 tokens and a maximum output of 8,192 tokens. The supported request parameters are include_reasoning, max_tokens, reasoning, seed, temperature and top_p. The reasoning and include_reasoning parameters are notable for a moderation model, since they let you control whether reasoning output is requested and whether it is included in the response.

Practical uses follow directly from those capabilities. You can run it as a pre-filter on user prompts before they reach a chat or generation model, and as a post-filter on model responses before they are shown to a user. Because it accepts images as well as text, the same model can screen multimodal traffic where a picture accompanies a request. The long context window means a full conversation or a lengthy document can be checked in one call instead of being split into chunks. Teams building assistants, community platforms or internal tools can use it as a consistent policy layer in front of several different generation models.

Calling it through api14 works the same way as any other model on the platform, using the OpenAI-compatible API with the model id nvidia/nemotron-3.5-content-safety:free. You send your request to the chat completions endpoint with that id in the model field, and you can set any of the supported parameters: include_reasoning, max_tokens, reasoning, seed, temperature and top_p. Because the interface is OpenAI-compatible, existing clients and SDKs generally need only the base URL and model id changed. The max_tokens value is bounded by the model's 8,192-token maximum output.

Keep in mind what this model is for. It is a 4B-parameter guardrail model, not a general-purpose assistant, so tasks like open-ended writing, complex reasoning or code generation are better served by a generation model, with Nemotron 3.5 Content Safety applied around it. Its output modality is text only, so it does not generate images even though it can read them.