IBM Granite Guardian 4.1 8B

Granite Guardian 4.1 8B is a judge, not an assistant. It is fine-tuned from Granite 4.1 8B — the same 40 layers, the same 128K window, the same vocabulary — and then trained to do exactly one thing: read a prompt or an answer and return yes or no against a stated criterion, wrapped in score tags. IBM is explicit that using it any other way is outside its intended purpose and may produce unsafe output. What it checks out of the box: general harm, social bias, profanity, violence, sexual content, unethical behaviour, jailbreak attempts, whether an answer is grounded in the supplied documents, whether the retrieved passage was relevant, and whether an agent's function call actually matches the tools it was given. The last two matter for anyone running retrieval or agent systems, where the failure mode is not obscenity but a confident invented citation. The change in version 4.1 is that the list stopped being fixed. Bring Your Own Criteria lets the operator write a rule in plain language — a format requirement, a length limit, a domain policy — and have the model judge against it. IBM's numbers for that are the largest jump in the release: multi-constraint IFEval accuracy rises from 0.458 to 0.844, and InfoBench scored by humans from 0.535 to 0.706. In practice this turns a safety filter into a general-purpose grader. It is also a hybrid thinking model. In thinking mode it produces a reasoning trace before the verdict, which costs time and tokens and buys accuracy on groundedness — 0.76 average balanced accuracy. In non-thinking mode it returns the verdict alone, which is what a production filter needs when it sits in front of every request. Safety on out-of-distribution material holds at 0.79 F1 without thinking; function-call hallucination detection improves from 0.74 to 0.79. One number is worth reading carefully. Used as a reward model for picking the best of several candidate answers, it scores 70.29 on the JETTS benchmark, ahead of every reward model IBM tested up to 70 billion parameters. That is a claim about a narrow task, made by the manufacturer, with no independent replication — but it points at what a small judge model is for. The published weights count 8.38 billion parameters against 8.79 billion for the Granite 4.1 8B it was tuned from. The architectures are identical; the difference is bookkeeping, because the base repository stores the shared embedding matrix twice. Licence is plain Apache 2.0. English only, and IBM warns the model can itself be attacked by adversarial input.

▸Videos

No videos yet.

▸Reviews