Meta releases Llama Guard 4 on Hugging Face Hub with Llama Prompt Guard 2 for multimodal safety

AI Impact Summary

Meta has released Llama Guard 4, a dense 12B multimodal safety model designed to detect 14 hazard categories across text and image inputs, plus the 86M/22M Llama Prompt Guard 2 classifiers for prompt injections. The models can run on a single 24 GB GPU and are configurable to enable or exclude categories, supporting flexible moderation pipelines before or after model generation. This expands production safety tooling on Hugging Face Hub, enabling teams to integrate stronger input/output filtering into LLM-based apps and image-text workflows, but it introduces new configuration choices and validation requirements to avoid over-blocking or missed threats.