- Home
- AI Content Moderation Software
- Llama Guard
Llama Guard
What it is
Llama Guard classifies 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 this task, covering 14 hazard categories: the 13 MLCommons categories and Code Interpreter Abuse. It supports English, French, German, Hindi, Italian, Portuguese, Spanish, and Thai, and was optimized to assess search and code interpreter tool calls. The model is free and open source for self-hosted deployment. Llama Recipes documentation explains configuration and customization, while a Llama Cookbook example uses Hugging Face Transformers to check prompts and model outputs; that demo requires access to the weights on Hugging Face. An int8 version reduces checkpoint size by about 40%, with very small impact on performance, according to the model card. Meta recommends deploying it alongside Llama 3.1, but it may increase refusals of benign prompts. The model card also warns of limits related to pretraining data and possible vulnerability to adversarial or prompt injection attacks. Some categories may need current factual knowledge for accurate evaluation.
For and against
For
- Free and open source for self-hosted use
- Covers 14 hazard categories
- Supports eight listed languages
- Can assess search and code interpreter tool calls
- Int8 version reduces checkpoint size by about 40%
Against
- May increase refusals of benign prompts
- May be vulnerable to adversarial or prompt injection attacks
- Some categories may require up-to-date factual knowledge
The verdict
Llama Guard offers self-hosted classification across a range of content categories, languages, and tool calls. Readers should account for its noted evaluation and attack limitations, as well as the possibility of benign prompts being refused.
Who it suits
Llama Guard suits teams deploying self-hosted content safety classification for LLM inputs, responses, and supported tool calls. Its supported languages and configurable policies may fit teams with varied moderation needs.
Plans and prices
Llama Guard 3-8B
8B parameters · model weights required
Read 4 Oct 2026 · github.com
On the record
- Customization
- Llama Recipes documentation explains how to configure and customize Llama Guard 3.github.com · read 4 Oct 2026
- 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.github.com · read 4 Oct 2026
- Hazard categories
- Llama Guard 3-8B covers 14 categories, including the 13 MLCommons hazard categories and Code Interpreter Abuse.github.com · read 4 Oct 2026
- 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.github.com · read 4 Oct 2026
- Languages
- It supports content safety classification in English, French, German, Hindi, Italian, Portuguese, Spanish, and Thai.github.com · read 4 Oct 2026
- 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.github.com · read 4 Oct 2026
- Model
- Llama Guard 3-8B is a Llama 3.1 8B pretrained model fine-tuned for content safety classification.github.com · read 4 Oct 2026
- Purpose
- Llama Guard classifies content in LLM inputs and responses as safe or unsafe.github.com · read 4 Oct 2026
- Quantization
- An int8 quantized version reduces checkpoint size by about 40% with very small impact on model performance, according to the model card.github.com · read 4 Oct 2026
- Recommended use
- Meta recommends deploying Llama Guard 3 alongside Llama 3.1, and says it may increase refusals of benign prompts.github.com · read 4 Oct 2026
- Tool use
- It was optimized for safety and security classification of search and code interpreter tool calls.github.com · read 4 Oct 2026
- Weights access
- The Cookbook demo lists access to Llama Guard model weights on Hugging Face as a requirement.github.com · read 4 Oct 2026
