Nemotron-Content-Safety-3.5

Nemotron Content Safety 3.5 is NVIDIA's June 2026 content moderation model, a 4 billion parameter fine-tune of Google's Gemma 3 4B IT that was trained as a LoRA adapter and merged back into the base weights. It takes a user prompt, an optional single image and an optional assistant response, and returns a structured verdict marking the user turn and the response as safe or unsafe together with the violated categories, using the same taxonomy as NVIDIA's Aegis Content Safety Dataset V2. The 128K context window is inherited from Gemma 3, and training and support cover twelve languages, including English, German, French, Spanish, Japanese, Korean, Arabic, Hindi and Chinese. The change from Nemotron 3 Content Safety is policy adaptation: a caller can pass its own moderation policy in natural language at inference time, suppressing default categories or adding domain-specific ones, and a thinking mode emits a reasoning trace before the verdict for auditing and policy iteration. NVIDIA reports harmful-content F1 of 0.86 on Aegis 2.0 prompts, 0.85 on XSTest prompts, 0.77 on WildGuard responses and 0.97 on Aya Redteaming. The weights are published under the OpenMDW-1.1 license, the Linux Foundation license NVIDIA has adopted for its open models, with Google's Gemma terms of use applying, and NVIDIA releases the training and evaluation dataset alongside them. The model is intended as a guardrail layer in front of and behind text and vision-language models rather than as a general assistant.