# Nemotron-3.5-Content-Safety-4B ![Organization](https://img.shields.io/badge/Org-NVIDIA-555) ![Pricing](https://img.shields.io/badge/Price-Open%20Weights-2da44e) ![License](https://img.shields.io/badge/License-NVIDIA%20Open%20Model-d29922) ![Modality](https://img.shields.io/badge/Modality-Text-8957e5) > **Quick answer:** Nemotron-3.5-Content-Safety-4B is NVIDIA's 4B open-weights safety classifier, trained on the Nemotron Content Safety Dataset V2 (formerly Aegis v2). It's the standout performer on multilingual toxicity (RTP-LX Prompt) in Mistral's Shieldstral comparison, leading all other models by a wide margin there. ## At a Glance **Where it leads:** RTP-LX Prompt (multilingual toxicity classification) — the clear leader among all comparison models. **Where it lags:** Refusal-detection and multilingual PolyGuard tasks relative to purpose-built classifiers for those specific tasks. **Best for:** Teams prioritizing multilingual toxicity detection in a small (4B), self-hostable footprint. ## What Nemotron-3.5-Content-Safety-4B Is Nemotron-3.5-Content-Safety-4B is NVIDIA's compact guard model trained on the Nemotron Content Safety Dataset V2 (Aegis v2), NVIDIA's 12-category, 9-subcategory hazard taxonomy dataset. It's part of NVIDIA's broader NeMo Guardrails ecosystem for content moderation, topic control, and jailbreak detection. ## Specifications | Field | Value | |---|---| | Organization | NVIDIA | | Parameters | 4B | | Training data | Nemotron Content Safety Dataset V2 (Aegis v2) | | License | NVIDIA Open Model License | | Modality | Text only | ## Pricing Open weights, free to download. ## Public Benchmark Scores *Scores from Mistral AI's Shieldstral model card, shown for context. Not Benchgen measurements.* | Benchmark | Score | |---|---| | [RTP-LX Prompt](/benchmarks/microsoft/rtp-lx-prompt) | 86.1% | | [WildGuardTest (Prompt)](/benchmarks/allen-institute-for-ai/wildguardtest-prompt) | 84.4% | | [Aegis v2 (Response)](/benchmarks/nvidia/aegis-v2-response) | 80.9% | | [RTP-LX Completion](/benchmarks/microsoft/rtp-lx-completion) | 89.7% | | [HarmBench (Prompt)](/benchmarks/center-for-ai-safety/harmbench-prompt) | 91.7% | ## Related Models & Benchmarks - [Shieldstral 1.0](/models/mistral-ai/shieldstral-1-0) · [Qwen3Guard-8B](/models/alibaba/qwen3guard-8b) - [Aegis v2 (Prompt)](/benchmarks/nvidia/aegis-v2-prompt) · [RTP-LX Prompt](/benchmarks/microsoft/rtp-lx-prompt) ## Frequently Asked Questions

What is Nemotron-3.5-Content-Safety-4B?

NVIDIA's 4B open-weights safety classifier trained on the Nemotron Content Safety Dataset V2 (Aegis v2), part of the NeMo Guardrails ecosystem.

Is it multimodal?

No, it's text-only.

Is it open source?
Yes, released under the NVIDIA Open Model License.

--- *Scores sourced from Mistral AI's [Shieldstral model card](https://huggingface.co/mistralai/Shieldstral-1.0-3B), shown for context. Last updated 2026-08-12.*

Nemotron-3.5-Content-Safety-4B — Training jobs

Training jobs