NVIDIA releases Nemotron 3 Content Safety, a 4B moderation classifier on Gemma-3

NVIDIA's model card for Nemotron 3 Content Safety gives a release date of 16 March 2026 (version V1.1) and says training ran from October 2025 to March 2026. It is a 4B-parameter LLM classifier that moderates prompts (text plus an optional image) and optional model responses. It returns safe or unsafe and, on request, the violated categories.

The base is Google Gemma-3-4B-it, fine-tuned with LoRA and merged back. The licence is the NVIDIA Nemotron Open Model License plus the Gemma Terms of Use and Prohibited Use Policy, and the card says it is ready for commercial use. It covers 12 languages and the output uses an Aegis-based taxonomy of 23 safety categories. It accepts one image, resized to 896x896.

From a partly garbled table the card reports Multijail prompt accuracy 0.92 with harmful F1 0.96, and Aegis 2 prompt accuracy 0.85 with harmful F1 0.87. False-positive rates on general benchmarks are 0.023 on MMMU and 0.058 on DocVQA, which the card's limits note as some over-flagging on document inputs.

Testing was on NVIDIA H100, A100 and RTX PRO 6000 under Linux. The card says use-case-specific testing is needed before deployment.

Kept here for context ahead of 3.5. Same shape, same advice to test on your own data.

Sources

• https://huggingface.co/nvidia/Nemotron-3-Content-Safety

Editor's first pass (AI-assisted)

• Real 4.5: Open weights with a commercial-use statement on the card.
• Useful 3.5: Prompt and response moderation with category output; some document over-flagging.
• Notable 3: Part of NVIDIA's Nemotron safety line.
• Novel 2.5: Follows the earlier Nemotron safety models.