Model reference · open weights

Llama-Guard-4

Llama-Guard-4 is an open-weight language model from Meta. Llama-Guard-4-12B (BF16) weighs 24.0 GB; the smallest configuration that runs it is RTX 5090 32 GB.

• Llama-Guard-4 is a 12-billion parameter multimodal safety classifier developed by Meta for classifying text and image content as safe or unsafe.
• It supports English and multilingual text prompts along with mixed text-and-image inputs, and it is licensed under an other license.
• The model uses a dense architecture pruned from Llama 4 Scout and is designed to run on a single GPU.

Summary of the meta-llama/Llama-Guard-4-12B model card, 2026-10-01

Hugging Face ↗GitHub ↗Paper ↗