Meta: Llama Guard 4 12B - Safety Classification Model

Ensure content safety with Llama Guard 4's multimodal capabilities.

Input: image · text Output: text Context: 163,840 tokens Release: 2025-04-30

Meta's Llama Guard 4 12B, released on April 30, 2025, is a versatile multimodal model designed for content safety classification. It is derived from Llama 4 Scout, offering a broad context window of 163,840 tokens. This model can process both text and images, generating text outputs that assess the safety of prompts and responses. Llama Guard 4 is aligned with the MLCommons hazards taxonomy and supports multiple languages, making it a robust tool for moderating diverse content types. Integrated with the Llama Moderations API, it effectively manages mixed text-and-image inputs, ensuring comprehensive content moderation.

Use Cases

Here are a few ways teams apply Meta: Llama Guard 4 12B in practice—from fast drafting to multimodal understanding. Adapt these ideas to your workflow.

Classify content safety in multilingual settings

Moderate mixed text and image content

Enhance content safety in digital platforms

Identify unsafe content categories

Key Features

A quick look at the capabilities that make this model useful in real projects.

Multimodal content classification

Supports text and image inputs

Aligned with MLCommons hazards taxonomy

Handles mixed text-and-image prompts

Integrated with Llama Moderations API

Specs

Overview

Vendor
meta-llama

Model ID
meta-llama/llama-guard-4-12b

Release
2025-04-30

Modalities & context

Input
image · text

Output
text

Context
163,840 tokens

Parameters & defaults

Supported parameters: frequency_penalty, logit_bias, max_tokens, min_p, presence_penalty, repetition_penalty, response_format, seed, stop, temperature, top_k, top_p

Defaults: temperature 0.2, top_p 0.95

Benchmark tests: Meta: Llama Guard 4 12B

We ran this model against a few representative prompts to show its range. Review the outputs below and be the judge.

Text

Prompt:

Write 150 words on how AI might positively upend work, leisure and creativity

safe

Run this prompt on Upend.AI

Image

Prompt:

Generate an SVG of a pelican riding a bicycle.

Run this prompt on Upend.AI

Ready to try Meta: Llama Guard 4 12B?

Chat with Meta: Llama Guard 4 12B

up.end

/ˌəpˈend/

verb

To “upend” means to completely disrupt, overturn, or drastically change the established order or structure of something. It implies a significant shift or alteration that can potentially have far-reaching consequences. When something is upended, it is turned upside down or transformed in a way that challenges conventional norms or expectations. The term often carries a sense of innovation, transformation, and sometimes even a hint of upheaval, indicating that the changes are not just minor adjustments but rather a fundamental reimagining of the status quo.

Privacy Terms