Meta Releases Llama Guard 4 for Multimodal Content Safety

Meta has released Llama Guard 4, a 12-billion parameter content safety classifier designed to moderate both text and image content in LLM applications.

Model Specifications

Llama Guard 4 is derived from Meta's Llama 4 Scout model and features a 163,840 token context window. The model is available free through OpenRouter with $0 per million tokens for both input and output. Its knowledge cutoff date is August 31, 2024.

Key Capabilities

The model classifies content safety in two modes: prompt classification (analyzing user inputs) and response classification (analyzing LLM outputs). According to Meta, it generates text output indicating whether content is safe or unsafe, listing specific violated content categories when unsafe content is detected.

Llama Guard 4 is aligned to the standardized MLCommons hazards taxonomy. The model supports English and multiple additional languages, though specific language lists were not disclosed.

Multimodal Features

The primary advancement over previous Llama Guard versions is multimodal capability. Llama Guard 4 can process mixed text-and-image prompts, including multiple images in a single request. This positions it as Meta's first safety classifier capable of handling the full range of inputs supported by multimodal Llama 4 models.

Integration and Availability

Meta has integrated Llama Guard 4 into the Llama Moderations API, providing safety classification for both text and images. The model is currently available through OpenRouter's routing infrastructure, which directs requests to providers based on prompt size and parameters.

Model weights are accessible, though specific hosting and licensing details were not provided in the release information.

What This Means

Llama Guard 4 addresses a critical gap in AI safety tooling by providing multimodal content moderation at no cost. As LLMs increasingly handle image inputs alongside text, safety systems must match these capabilities. The 164K context window is particularly relevant for applications that need to classify long conversations or multiple images simultaneously. Meta's alignment to the MLCommons taxonomy provides standardization that could improve interoperability across different AI safety systems.