NVIDIA releases Nemotron 3.5 Content Safety, a 4B multilingual moderation classifier
Gemma-3-4B-based, 12 languages, text plus one image, with custom-policy mode. Open licence plus Gemma terms; card dated 2 June 2026.
NVIDIA's Nemotron 3.5 Content Safety model card is dated 2 June 2026 (version V1.2). It classifies prompts, an optional image and optional responses as safe or unsafe. It has about 4B parameters, is based on Google Gemma-3-4B-it with a LoRA adapter merged back, and uses a SigLIP vision encoder at 896x896.
Standard mode follows the Aegis Content Safety Dataset V2 taxonomy and outputs a list of violated categories. A custom-policy mode accepts user-defined policies and can emit a reasoning trace first. It covers 12 languages: English, Arabic, German, Spanish, French, Hindi, Japanese, Thai, Dutch, Italian, Korean and Chinese.
The licence is OpenMDW 1.1 plus the Gemma Terms of Use and Prohibited Use Policy. The card says it is ready for commercial use. Context is up to 128K tokens, but multimodal input is limited to text plus one image. It is designed for NVIDIA GPUs on Linux.
The card's table includes VLGUARD prompt accuracy 0.89, Aegis 2.0 prompt F1 0.86, XSTEST response F1 0.87 and PolyGuard prompt 0.80. The page's table is hard to parse, so these are partial figures. On MM-SafetyBench accuracy is 0.55. False-positive rates on safe benchmarks are low, with DocVQA at 0.060.
The card says deployments need additional testing with use-case-specific data.
A maintained guard family with an honest MM-SafetyBench row.
Sources
Editor's first pass (AI-assisted)
- Real 4.5: Open weights on the Hub under OpenMDW 1.1 plus Gemma terms; the card says commercial use is ready.
- Useful 4: Flags prompts and responses, in 12 languages, with an optional reasoning trace.
- Notable 3.5: An incremental step in a maintained family from a major vendor.
- Novel 3: A version bump on the March 2026 release.
Comments