Llama Prompt Guard 2 22M is a compact classifier from Meta's Purple Llama initiative, designed to screen incoming text for prompt injection and jailbreak attempts before they reach a downstream LLM. Its small 22M-parameter footprint is the central design choice: by keeping the model tiny, it can run inline on every request with minimal added latency, and Groq positions it as a low-overhead guard layer that aims to cut latency and compute cost versus larger safety classifiers. Groq's documentation describes it as a specialized, real-time defense suited for production traffic where speed and predictability matter more than general reasoning ability.

In practical use, the model fits well as a first-pass filter in agentic or user-facing pipelines: classify the prompt, route clean traffic through the main LLM, and either rewrite or block suspicious inputs. Its narrow focus on prompt attacks, rather than broad content moderation, makes it a complementary building block rather than a standalone assistant. Pairing it with a larger reasoning model lets teams keep the heavyweight model shielded by a fast, cheap guard, which is the architecture Purple Llama and similar safety stacks increasingly recommend for LLM applications.