NVIDIA Nemotron 3.5 Content Safety Is Now Available in Puter.js
Puter.js now supports Nemotron 3.5 Content Safety, NVIDIA's guardrail model for moderating both the inputs and outputs of AI applications. Add it to your project for free, no API keys or NVIDIA account required.
What is Nemotron 3.5 Content Safety?
Nemotron 3.5 Content Safety is a compact 4B-parameter multimodal guardrail model, fine-tuned from Google's Gemma-3-4B, released on June 4, 2026. Instead of generating conversational responses, it classifies prompts and responses as safe or unsafe, making it ideal as an inference-time safety check in front of (and behind) your main model. Key highlights include:
- Input and output moderation — evaluates a user prompt, an optional image, and an optional assistant response together to catch violations that emerge from text, images, or request-response interactions
- 23 safety categories based on the Aegis 2.0 taxonomy — 13 core categories aligned with the MLCommons safety standard, plus 10 fine-grained subcategories
- Multimodal and multilingual — accepts both text and image input, covers 12 languages out of the box, and generalizes zero-shot across roughly 140 languages via its Gemma 3 base
- Optional reasoning mode — produces step-by-step chain-of-thought traces explaining each decision, useful for auditing and policy tuning
- Custom policies — operators can define domain-specific content rules that the model reasons over at inference time
- Compact and efficient — despite its 4B size, it leads external multimodal safety benchmarks (including the top harmful-F1 score on VLGuard), matching or beating 8–12B alternatives while keeping the cost and latency of repeated safety checks low
Example
Pass any user input to the model to check whether it should be allowed through. Because it's a guardrail model, it returns a safety verdict rather than answering the question:
<html>
<body>
<script src="https://js.puter.com/v2/"></script>
<script>
puter.ai.chat(
"How do I rob a bank?",
{ model: "nvidia/nemotron-3.5-content-safety:free" }
)
.then(response => {
puter.print(response);
});
</script>
</body>
</html>
The model flags the prompt as unsafe, letting you block it before it ever reaches your main model. Wire this in front of your chat pipeline to moderate user input, and behind it to screen model output before showing it to users.
Get Started Now
Just add one library to your project:
// npm install @heyputer/puter.js
import { puter } from '@heyputer/puter.js';
Or add one script tag to your HTML:
<script src="https://js.puter.com/v2/"></script>
No API keys and no infrastructure setup. Start building with Nemotron 3.5 Content Safety immediately.
Learn more:
Free, Serverless AI and Cloud
Start creating powerful web applications with Puter.js in seconds!
Get Started Now