Llama Guard is a series of input and output moderation models from Meta, part of the Purple Llama project. The models are meant to screen what a language model receives and returns and to recognise content that violates usage guidelines. Llama Guard 3 is fine-tuned from Llama 3.1 and 3.2 and detects hazards according to the MLCommons taxonomy; versions exist at 8B, 1B and an 11B model with vision support. Developers can adapt the models to their own use case. The weights are distributed through Hugging Face and you run them yourself, for example as part of your own AI pipeline.
Meta is a US company and controls the models and their licence terms. The models are released under the Llama Community License; Meta describes Purple Llama components as permissively licensed, but this is a bespoke licence rather than a standard open-source licence. Because you can run the models in your own environment, prompts and responses need not reach Meta. You do remain dependent on Meta for updates and for any change in licence terms.