Granite Guardian vs Llama Guard 4

Granite Guardian

One ungated Apache 2.0 model judges harm, jailbreaks, RAG groundedness, function-call errors and custom criteria, with signed weights and published evaluation code. It is trained and tested on English only, each call checks one criterion, the 4.1 prompt format differs from 3.x, and IBM's watsonx.ai lists only the deprecated 3.0 model.

Llama Guard 4

A single self-hosted model classifies text and multi-image prompts against 14 MLCommons-aligned hazard categories and answers in a few tokens. The weights have not changed since 29 April 2025, download access needs Meta's manual approval, and the licence withholds the grant from individuals and companies based in the European Union.

Granite Guardian

1. Append the <guardian> block as the final user message, with the mode line, ### Criteria: and ### Scoring Schema:. Copy the strings from the model card, because no package builds them
2. Use the no-think instruction for gating and parse <score>. Think mode writes a reasoning trace first, and the card's examples allow up to 2,048 output tokens
3. Treat yes as the criterion being met, which for built-in criteria means the risk is present. Treat a missing <score> tag as a failed check
4. Pass retrieved text through documents= and tool schemas through available_tools= in apply_chat_template, not inside the message text
5. Under Ollama, set num_ctx in the request options. IBM's docs say the default context is short and long requests are truncated

Llama Guard 4

1. Request access on the Hugging Face page before anything else. Approval is manual, and the form cannot be edited after submission
2. Send only the user turn to check an input, and the user turn plus the model's answer to check an output. The template picks the role from the message count
3. Parse the first line for safe or unsafe and the second for category codes. Set max_new_tokens to about 10 and turn sampling off
4. Do not send an image with no text. Meta says the model is not an image-only classifier, and S14 is skipped when an image is present
5. Pair it with a prompt-attack detector. The card says the model can itself be moved by adversarial or injected text