Nemotron 3.5 Content Safety

Nemotron 3.5 Content Safety is a compact 4B-parameter multimodal guardrail model published by NVIDIA, fine-tuned from Google Gemma-3-4B. It is designed to moderate both inputs to and responses from large language models and vision-language models, accepting text and image input while returning text output. Each inference produces a safe or unsafe classification for the user prompt and the assistant response, along with safety category labels and an optional reasoning trace. The model covers 12 languages, making it suitable for multilingual deployment in production safety pipelines.

The model became publicly available on June 4, 2026, going live on the DeepInfra inference platform on the same day as the broader Nemotron 3 Ultra and Nemotron ASR releases. Weights are openly published on Hugging Face under the nvidia namespace, allowing organizations to self-host the guardrail or run it through hosted gateways. Its small footprint and multimodal coverage make it a practical choice for teams that need lightweight, open-weight content moderation alongside a primary generative model, rather than relying on a large general-purpose model for safety classification.

Requestynemotron-3.5-content-safetynemotron

Open provider page Provider docs

Quick Info

Provider: Requesty
Model key: nemotron-3.5-content-safety
Release date: Jun 4, 2026
Last updated: Jun 4, 2026
Input modalities
Output modalities
Capabilities

Cost

A provider subscription or plan supersedes token-based pricing for this model.

Limits

Output tokens: 8,192 tokens
Context window: 131,072 tokens

Latest news about Nemotron 3.5 Content Safety

Requesty

Coverage

Kingy AI published a technical guide on NVIDIA Nemotron 3.5 Content Safety, describing it as a 4-billion-parameter model released publicly in June 2026 for classifying text, a single image, and an optional model response under either NVIDIA's standard safety taxonomy or a developer-supplied policy. The guide positions Citing the official model card, the Kingy AI guide lists a Gemma 3 4B base with a 128K context window and a SigLIP vision encoder, with explicit support for 12 languages and broader zero-shot language coverage from the Gemma base. The article confirms the model accepts a user prompt, optional image, optional assistant

Requesty

Coverage

Vultr announced Day Zero support for NVIDIA Nemotron 3.5 Content Safety, a compact 4-billion-parameter multimodal safety model designed to moderate AI inputs and outputs across text and images. According to the Vultr blog, the model is built on Google's Gemma-3-4B-it foundation and fine-tuned by NVIDIA, combining multi The Vultr post details key capabilities of Nemotron 3.5 Content Safety, including 23 safety categories based on the Aegis v2 taxonomy, custom-policy enforcement with reasoning-based explanations, reasoning-based moderation workflows, multilingual moderation for 12 languages out of the box, multimodal text-plus-image ev