Meta releases Llama Guard 4 (12B multimodal safety model) and Llama Prompt Guard 2 on Hugging Face Hub

AI Impact Summary

Meta released Llama Guard 4, a 12B dense multimodal safety model, along with two Llama Prompt Guard 2 classifiers, and hosts open checkpoints on Hugging Face Hub. The model can evaluate both text and image inputs and classify 14 hazard categories (S1–S14) with configurable exclusions, enabling pre-filtering and post-generation moderation in production pipelines. It operates as a dense non-MoE architecture that can run on a single 24 GB GPU, with example code showing integration through transformers AutoProcessor and Llama4ForConditionalGeneration to support multilingual safety checks.