LlamaGuard Now Available on the CentML Platform

CentML welcomes three new members to the Llama herd

Today at LlamaCon, Meta announced the release of 3 new members of the Llama family of open-source models: Llama-Guard-4-12B, Llama-Prompt-Guard-2-22M, and Llama-Prompt-Guard-2-86M. As a launch partner, CentML is excited to offer these models to customers on their day of release.

Llama-Guard-4-12B

The model is part of the Llama Guard series designed for input-output safeguards in human-AI conversations, and utilizes instruction tuning methodology to follow specific guidelines and adapt to different taxonomies. A natively multimodal safety classifier with 12 billion parameters,  Llama Guard 4 was trained jointly on text and multiple images and fine-tuned on specially curated datasets for safety risk classification. Llama-Guard-4-12B is a lightweight LLM designed for use in content moderation for conversational AI systems and provides safety enhancement in applications like chatbots and virtual assistants. 

According to the model card, the model is trained to predict safety labels based on the MLCommons safety taxonomy, and includes an additional category, Code Interpreter Abuse, for text-only tool-call use cases involving phishing and other malicious activities. Llama Guard 4’s outputs indicate whether a given prompt or response is safe or unsafe, and also lists any content categories that were violated.

Llama-Prompt-Guard-2-22M

Designed to detect and prevent malicious prompt attacks on LLM-powered applications, this model was built as a multi-label classifier using mDeBERTa-v3-base (a multilingual version of DeBERTa). It features a 512-token context window for efficient processing and classifies inputs into three categories: benign, injection, and jailbreak. Depending on the use case, it may require fine-tuning for optimal performance.

Llama-Prompt-Guard-2-86M

Although limited information is available about this model, we anticipate it will be used to enhance security and safety in AI systems, provide real-time content moderation, and evaluate quality in AI pipelines. It was also built using mDeBERTa-v3-base.

CentML Platform

The CentML Platform gives users secure access to popular open-source models like Llama, either through a serverless API or dedicated endpoints. CentML optimizes all layers of the AI stack — from application to silicon — delivering peak performance without sacrificing infrastructure flexibility. The platform lets users easily switch between accelerators, cloud vendors, and models and customize their deployments for the optimal balance of price and performance. It supports all popular AI applications (RAG, chatbots, agentic, vision) and offers the ability to finetune models with custom data for more accurate results.

Give the new LlamaGuard models a try on our Platform today! 

• Llama-Guard-4-12B is available through our Serverless API
• Llama-Prompt-Guard-2-22Mand Llama-Prompt-Guard-2-86M are available through our General Inference