Llama Guard 4 12B

Llama Guard 4 12B is a multimodal safety classifier from Meta that screens both user prompts and model outputs for harmful content. It is a 12-billion-parameter model derived by pruning the Llama 4 Scout foundation and keeping only the shared expert, which yields a dense feedforward early-fusion design rather than a mixture-of-experts layout. This pruned architecture is intentionally compact enough to run on a single GPU with 24 GB of VRAM, making it practical for real-time guardrail filtering alongside larger Llama 4 deployments. The model evaluates content against the MLCommons hazard taxonomy along with a code interpreter abuse category, covering areas such as violent crimes, sexual content, hate speech, self-harm, and intellectual property concerns.

As a multimodal classifier, Llama Guard 4 handles multilingual text and mixed text-and-image prompts, including multi-image inputs, unifying the text-only coverage of the earlier Llama Guard 3-8B and the vision capability of Llama Guard 3-11B-vision into a single model. Meta reports meaningful gains over Llama Guard 3, with roughly 4 percent higher English recall, about 8 percent better F1, and 17 to 20 percent improvements on multi-image detection. These advances make it well suited as a lightweight safety layer for multimodal Llama 4 applications, where consistent classification across language and image content is more important than generative capability.