Llama Guard
ai.meta.com
- Free plan
- Self-hosted
What Llama Guard is
Llama Guard is a free, open-source model for classifying content in large language model inputs and responses as safe or unsafe. Llama Guard 3-8B is a Llama 3.1 8B pretrained model fine-tuned for content-safety classification. It covers 14 hazard categories, including the 13 MLCommons categories and Code Interpreter Abuse, and supports English, French, German, Hindi, Italian, Portuguese, Spanish, and Thai. The model is optimized for safety and security checks on search and code-interpreter tool calls. Llama Recipes explains configuration and customization, while Meta’s Llama Cookbook demonstrates checking prompts and model outputs with Hugging Face Transformers. That example requires access to weights on Hugging Face. An int8 version reduces checkpoint size by about 40% with very small reported performance impact. Meta recommends deploying it with Llama 3.1, but notes it may refuse benign prompts and can be vulnerable to adversarial or prompt-injection attacks.
Who it suits
Llama Guard suits teams deploying their own content-safety checks around language-model prompts, responses, search, or code-interpreter calls. It fits multilingual self-hosted systems that can obtain and run the model weights.
Good to know, and what to watch
- Free open-source model
- Covers 14 hazard categories
- Supports eight listed languages
- Customizable through Llama Recipes
- Checks prompts and model outputs
- Requires self-hosted deployment
- Model weights access is required
- May refuse some benign prompts
- Can face prompt-injection vulnerabilities
Plans
Every plan and priceLlama Guard 3-8B
Open source
The verdict
Llama Guard provides a multilingual safety-classification layer for self-hosted LLM systems. Its documented caveats around refusals, adversarial inputs, and factual knowledge should shape deployment decisions.
What the maker publishes
- Customization
- Llama Recipes documentation explains how to configure and customize Llama Guard 3. source
- Evaluation caveat
- The model card notes that some categories, including defamation, intellectual property, and elections, may require factual, up-to-date knowledge for accurate evaluation. source
- Hazard categories
- Llama Guard 3-8B covers 14 categories, including the 13 MLCommons hazard categories and Code Interpreter Abuse. source
- Integration example
- Meta’s Llama Cookbook demonstrates using Llama Guard as an inference safety checker for both user prompts and model outputs with Hugging Face Transformers. source
- Languages
- It supports content safety classification in English, French, German, Hindi, Italian, Portuguese, Spanish, and Thai. source
- Limitations
- The model card says performance can be limited by pretraining data and that the model may be vulnerable to adversarial or prompt injection attacks. source
- Model
- Llama Guard 3-8B is a Llama 3.1 8B pretrained model fine-tuned for content safety classification. source
- Purpose
- Llama Guard classifies content in LLM inputs and responses as safe or unsafe. source
- Quantization
- An int8 quantized version reduces checkpoint size by about 40% with very small impact on model performance, according to the model card. source
- Recommended use
- Meta recommends deploying Llama Guard 3 alongside Llama 3.1, and says it may increase refusals of benign prompts. source
- Tool use
- It was optimized for safety and security classification of search and code interpreter tool calls. source
- Weights access
- The Cookbook demo lists access to Llama Guard model weights on Hugging Face as a requirement. source
Checked against the maker's pages on 4 Oct 2026.