Meta releases Llama Guard 4 and Llama Prompt Guard 2 on Hugging Face Hub

AI Impact Summary

Meta has released Llama Guard 4, a new 12B dense multimodal safety model, and two Llama Prompt Guard 2 models, designed to detect unsafe content in images and text generated by large language models. This release introduces a new model checkpoint available on the Hugging Face Hub, alongside an interactive notebook for easy experimentation. The new models offer improved performance and a faster 22M parameter model, along with tokenization resistant to adversarial attacks, and simplified binary classification for prompt injection detection.