Llama Prompt Guard 2 is a specialized 86M parameter classifier from Meta's Purple Llama initiative, built to detect and prevent prompt attacks in LLM applications. It identifies malicious inputs like prompt injections and jailbreaks across multiple languages, functioning as a dedicated guardrail rather than a general-purpose language model. The model operates as a classification head that categorizes inputs into risk tiers, and community discussions show users can fine-tune this head to expand from three to six classification categories.

The model prioritizes efficient, real-time protection while keeping latency and compute costs low. Running on Groq's infrastructure, it uses TruePoint Numerics quantization, which selectively reduces precision only in areas that do not affect accuracy—preserving detection quality while delivering faster inference. This combination of specialized safety training and hardware-optimized serving makes Prompt Guard 2 well-suited for production deployments where ongoing input filtering is needed without sacrificing response speed.