Llama Guard 4 is a 12-billion-parameter safety classification model that detects harmful content in both model inputs and outputs. It is built by pruning the Llama 4 Scout foundation, keeping only the shared expert to give it a dense feedforward early-fusion architecture rather than a mixture-of-experts design, which lets it run on a single GPU with 24GB of VRAM. The model classifies content against the hazard categories defined in the MLCommons taxonomy along with code interpreter abuse, covering areas such as violent crimes, sexual content, hate speech, self-harm, and intellectual property. It evaluates multilingual text and can process mixed text-and-image prompts, combining the capabilities of the earlier Llama Guard 3-8B and 3-11B-vision models into a single classifier with added multi-image support. Meta reports that, compared to Llama Guard 3, version 4 improves English recall by about 4% and F1 by about 8%, with roughly 17 to 20% gains on multi-image detection. It is designed to safeguard multimodal Llama 4 applications while staying light enough for real-time prompt and response filtering.
Llama Guard 4 12B runs on 1 route through the Opper gateway. Compare residency, ZDR, and training posture at a glance — full data-handling detail per route below.
| Provider | Region | Zero data retention | Training | Input | Output |
|---|---|---|---|---|---|
| US | Enterprise | No | $0.18 | $0.18 |
Training posture across routes: No training on prompts by default.
Each route hosting Llama Guard 4 12B has its own privacy posture, residency, and GDPR terms. Postures are maintained by Opper with a last-verification timestamp.
Zero data retention is available via Opper Enterprise contract. No training on customer data. US; unknown.
Call Llama Guard 4 12B through the Opper gateway with one API key. Let your coding agent set it up, or call it directly — Opper is drop-in compatible with the OpenAI, Anthropic, and Google AI SDKs.
Copy this and paste it into your coding agent — Claude Code, Cursor, Codex, and more — and it'll wire up Opper for you.
import OpenAI from "openai";const client = new OpenAI({apiKey: process.env.OPPER_API_KEY,baseURL: "https://api.opper.click/v3/compat",});const completion = await client.chat.completions.create({model: "deepinfra/meta-llama/Llama-Guard-4-12B",messages: [{ role: "user", content: "Hello" }],});console.log(completion.choices[0].message.content);
Side-by-side on privacy, EU hosting, pricing, and benchmarks.
