Llama Guard 4

Llama Guard 4, released April 2025 alongside Llama 4, is Meta's input-output safeguard model: a 12B dense classifier pruned from Llama 4 Scout and fine-tuned to judge whether prompts and responses are safe. It is natively multimodal — text, images, or mixed — and classifies against the standardized MLCommons taxonomy of 14 hazard categories (violent crimes, hate, self-harm, privacy, elections, code-interpreter abuse, and more), outputting safe or unsafe plus violated categories.

It consolidates the previous generation's split models (Llama Guard 3-8B for text, 3-11B-Vision for images) into one, adds multi-image prompt support and custom categories, and ships with a 163K-token context window. Weights are downloadable (gated) under the Llama 4 Community License, and it is integrated into Meta's Llama Moderations API.

Honest caveats: it is optimized for English, the license is Meta's custom community license (not OSI-open), and like all classifiers it can be jailbroken — it is one layer in a defense-in-depth stack, not a guarantee. For teams building on Llama 4, it is the natural first safety filter.