Nemotron 3.5 Content Safety
Neural Network
Nemotron 3.5 Content Safety is a 4B parameter NVIDIA safety model for moderating LLM requests and responses.
Specifications
- Input
- text, images, documents
- Output
- text
- Context size
- 128 000 tokens
- Max answer length
- 8 192 tokens
- Reasoning effort
- low, medium, high
- Capabilities
- reasoning
How it works Nemotron 3.5 Content Safety?
Nemotron 3.5 Content Safety is a compact 4B parameter NVIDIA safety model, fine-tuned based on Google Gemma-3-4B. Its task is not to generate content, but to verify it: the model classifies both incoming user requests and responses from language and vision-language models, working with text and images. Its compact size allows it to be used as a filter both before and after the main model. In BotHub, it is accessible without a VPN or foreign card, payment is made with Russian cards in rubles based on usage, and a unified OpenAI-compatible API allows you to connect content verification to your service and switch to any of over 250 models at any time without rewriting the integration. Traffic is encrypted via AES-GCM, chat history is not saved, and contracts, invoices, and electronic document management are available for companies. Use cases are clear: moderating chatbot messages, checking assistant responses before output, filtering uploaded images, labeling archives, and offloading manual moderation teams.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 model from the Nemotron family, focused on content analysis and safety. It accepts text, documents, and images, reasons before answering, and is suitable for message classification, user content verification, moderation scenarios, and analytical tasks with a large context of up to 128,000 tokens.
The model outputs up to 8192 tokens per response — enough for a detailed analysis, comprehensive conclusion, or structured report. If you need a longer result, break the task into parts: the 128,000-token context allows you to keep the entire conversation history.
Yes, the model's reasoning mode is confirmed. Before outputting, it builds a chain of steps, which helps to more accurately analyze controversial and edge cases, compare wording against criteria, and provide a reasoned conclusion rather than just a label. Responses to complex queries may take longer.
Our data does not indicate support for function calling or strict structured output, so you shouldn't rely on it in advance. In practice, you can describe the required format directly in the prompt and test the result. Requests go through the unified OpenAI-compatible BotHub API.
The model accepts images and documents as input — you can send a picture or file and get a text analysis. There is no information about audio or video in our data. If you need such formats, specialized models are available nearby in BotHub, and switching takes just one action.
NVIDIA Nemotron family models were trained on multilingual corpora, so it works with Russian texts. The exact quality depends on the task and wording, so test it with your own materials: in BotHub, this is done without a VPN or foreign card, with payment in rubles.