Nemotron 3.5 Safety: 정적 필터가 멀티모달 규모에서 실패하는 이유
NVIDIA가 12개 언어와 이미지-텍스트 추론을 지원하는 커스터마이즈 가능한 가드레일 모델 Nemotron 3.5 Content Safety를 공개하며, 멀티모달 엔터프라이즈 AI의 허점을 막는 것을 목표로 하고 있습니다.
nvidia released nemotron 3.5 content safety, a multimodal guardrail model built on the gemma 3 4b architecture that processes user prompts, images, and responses in a unified context window. unlike static classifiers, this model supports custom policy specifications for different industries and outputs auditable reasoning traces in its 'think' mode to explain safety verdicts. nvidia also open-sourced the accompanying multimodal, multilingual safety dataset, enabling developers to benchmark and train safety systems on restrictive content without licensing hurdles.