llama-guard-4-12b

Llama Guard 4 12B is Meta’s multimodal safety classification model for moderating text, images, and mixed text-image inputs. It can evaluate both incoming prompts and generated responses, returning safe or unsafe classifications along with the relevant policy categories when violations are detected. The model supports multilingual text and multiple images, making it well suited for chat moderation, multimodal AI safety, prompt and response filtering, policy enforcement, and agent guardrails. It is built on a dense 12B architecture derived from Llama 4 Scout and is designed as a dedicated safety layer for production AI applications.