Llama Guard 3

Released in July 2024 on the Llama 3.1 base, Llama Guard 3 is Meta's open safety classifier that labels prompts and model responses as safe or unsafe across the MLCommons hazard taxonomy, adding categories like code-interpreter abuse and elections over Guard 2. It ships in 8B, a 1B edge variant, and an 11B multimodal vision variant, and is widely used as a self-hosted moderation layer.

Limitations: binary safe/unsafe outputs with no confidence scores; primarily English-focused; Meta's Llama community license carries use restrictions; the line has seen no new generation since Llama Guard 4 in April 2025.