Llama Guard 4 12B is Meta's purpose-built safety classifier, derived by pruning the Llama 4 Scout foundation and retaining only the shared expert. This produces a 12-billion-parameter dense feedforward early-fusion network rather than a mixture-of-experts design, which is light enough to run on a single 24GB-VRAM GPU. That lineage is central to its identity: it consolidates the text coverage of the earlier Llama Guard 3-8B and the vision coverage of Llama Guard 3-11B-vision into one model, adding multi-image support so a single classifier can guard multimodal Llama 4 applications end to end.

In practice, the model evaluates multilingual text along with mixed text-and-image prompts, returning a text verdict that flags content as safe or unsafe and, when unsafe, lists the violated hazard categories drawn from the MLCommons taxonomy plus a code-interpreter-abuse category covering areas such as violent crimes, sexual content, hate, self-harm, and intellectual property. Meta reports measurable gains over the previous generation, including roughly four-percent higher English recall, about eight-percent better F1, and seventeen-to-twenty-percent improvements on multi-image detection, making it well suited to real-time prompt and response filtering in production LLM pipelines.