Llama-Guard-3-11B-Vision

Llama Guard 3 11B Vision is Meta's multimodal content-safety classifier, fine-tuned from Llama 3.2 11B Vision to safeguard human–AI conversations that involve image understanding. Unlike earlier text-only Llama Guard models, it classifies both multimodal prompts (text plus image) and text responses against the MLCommons hazard taxonomy across 13 categories, including violent crimes, hate, self-harm, child exploitation, elections, and indiscriminate weapons.

The model acts as an LLM-based guardrail: given a conversation, it outputs whether the target message is safe or unsafe and, if unsafe, lists the violated category codes. On Meta's internal benchmark it reaches an F1 of 0.938 for response classification and 0.733 for prompt classification, outperforming GPT-4o and GPT-4o mini with substantially lower false-positive rates. It is optimized for English, supports one image per request (rescaled into four 560×560 chunks), and is intended for deployment alongside vision-capable LLMs rather than as a general chat or reasoning model.

Llama Guard 3 11B Vision is released under the Llama 3.2 Community License with open weights on Hugging Face. It is not designed for standalone text-only or image-only moderation; Meta recommends pairing it with text-only guards such as Llama Guard 3 8B or 1B where appropriate. Performance may vary with image resolution, adversarial inputs, and categories requiring up-to-date factual knowledge such as defamation or elections.