Nemotron 3.5 Content Safety
Neural Network
Nemotron 3.5 Content Safety is a 4B parameter NVIDIA guardrail model for checking LLM and multimodal model requests and responses
Specifications
- Input
- text, images, documents
- Output
- text
- Context size
- 131 072 tokens
- Max answer length
- 117 964 tokens
- Reasoning effort
- low, medium, high
- Capabilities
- reasoning
Prompt cost
(per 1M tokens)
Answer cost
(per 1M tokens)
How it works Nemotron 3.5 Content Safety?
Nemotron 3.5 Content Safety is a compact 4B parameter NVIDIA model, fine-tuned on Google Gemma 3 4B, designed for content safety classification. It acts as a filter around other models: it checks both what the user sends to a language or multimodal model and what the model responds with. It accepts text and images, making it suitable for image-based chats. Due to its small size, this check can be applied to every request. Connection in BotHub is via a unified OpenAI-compatible API: the same key and format as other models in the catalog (over 250), so you can change the main model behind the filter without rewriting the integration. Access is open from Russia without VPN or foreign cards, payment in rubles upon usage, data is encrypted via AES-GCM and not stored, and for companies, there is a contract, invoice, EDI, and admin panel with limits. The model is useful for filtering requests in support and product bots, checking generative model responses before display, filtering out unwanted images in multimodal services, and labeling dialogue logs for audit.Frequently asked questions about Nemotron 3.5 Content Safety
You can use generated results for commercial purposes. You own all rights to the content you create. The only restriction: make sure your prompt does not include copyrighted third-party material. You are responsible for respecting the rights to any input data.
This is an NVIDIA text model with a 131,072 token context: it parses long documents, accepts images and files, writes and edits code, and answers questions based on materials. As the name suggests, the model's profile is content safety assessment and moderation, making it convenient to use as a filter in a pipeline.
The output limit is 117,964 tokens per response, which is tens of thousands of words. This capacity is enough for a detailed analysis of a large document, a comprehensive audit report, or a long block of code without splitting the request. The total dialogue budget is 131,072 tokens including your request.
Yes, reasoning mode is confirmed: the model first builds a train of thought and then provides a result. This significantly helps where context and edge cases need to be weighed — for example, deciding if a fragment violates rules and explaining the basis for such a verdict.
Our data does not indicate support for function calling or structured output, so we cannot guarantee it — please test it in your scenario. Often, explicitly specifying the format in the prompt helps. If your tools require strict JSON, you can switch to another model in BotHub via the same API without rewriting the integration.
Images and documents — yes, this is confirmed: you can send an image or file and ask to analyze the content, including for admissibility. Audio and video input are not noted in our data. The model responds with text — generation of images, audio, or video is not claimed.
NVIDIA Nemotron family models were trained on multilingual corpora, so the model works with Russian text, but we have no data on the quality specifically for it. Test it with your own examples: in BotHub, access is open from Russia without VPN or foreign cards, payment in rubles upon usage.