Meta Llama Guard 4 12B

Llama Guard 4 12B is a dense, 12-billion parameter multimodal safety classifier engineered to act as a robust guardrail for large language model applications. Its primary design intent is to evaluate both text and image inputs, providing a systematic way to classify content as safe or unsafe. When it identifies potentially harmful material, the model explicitly lists the violated content categories, allowing developers to implement precise moderation logic. By supporting mixed text-and-image prompts, it serves as a versatile tool for maintaining safety standards in complex, modern AI workflows.

Derived from the Llama 4 Scout architecture, this model benefits from a lineage focused on high-performance content classification and alignment with the standardized MLCommons hazards taxonomy. Through its fine-tuned training, it offers broad language support and enhanced capabilities for processing multiple images within a single prompt. Integrated into the Llama Moderations API, the model is built to provide consistent, scalable safety checks, making it a practical choice for teams looking to integrate reliable, automated moderation into their digital platforms and multimodal AI systems.