Nvidia releases an open multimodal safety model with custom policies

Nvidia released Nemotron 3.5 Content Safety, a compact 4-billion-parameter moderation model built on Gemma 3 that evaluates a prompt, optional image, and optional assistant response together. The model supports explicit training coverage for 12 languages, broad zero-shot multilingual transfer, natural-language custom safety policies, and an optional THINK mode that produces an auditable reasoning trace before its verdict. Nvidia also released a multimodal, multilingual safety dataset, an uncommon step in a field where licensing and sensitive-content constraints often keep training artifacts closed. Nvidia reports roughly 85% average accuracy across its evaluated multilingual and multimodal benchmark set and says the default mode retains low latency. Those are vendor-reported results, but the release is technically meaningful because it packages policy customization, multimodal moderation, and inspectable data into an open commercial-use model.

Key details: June 4, 2026, 4B parameters, Gemma 3 base, 12 explicitly trained languages, Approximately 140-language zero-shot reach, Open safety dataset.

Continue swiping for more AI Brief stories.