Llama Guard 4 12B

Llama Guard 4 12B is a 12-billion-parameter safety classification model engineered to filter harmful content in both user prompts and model responses, serving as a guardrail layer for Llama 4-based applications. Rather than using a mixture-of-experts design, it is created by pruning the Llama 4 Scout foundation and keeping only the shared expert, producing a dense feedforward early-fusion architecture that can run on a single 24GB GPU. This makes it practical for production deployments that need consistent moderation without heavy infrastructure overhead.

The classifier evaluates content against the MLCommons hazard taxonomy with added coverage for code interpreter abuse, spanning categories such as violent crimes, sexual content, hate speech, self-harm, and intellectual property. It handles multilingual text and can score mixed text-and-image prompts, combining the strengths of the earlier text-only and vision Llama Guard 3 variants into one model with multi-image support. Meta reports notable gains over Llama Guard 3, including roughly 4% better English recall, about 8% higher F1, and 17 to 20% improvements on multi-image detection, positioning it as a lightweight, real-time filter for multimodal assistants.