IBM releases Granite Guardian 4.1 8B, a judge model for safety, RAG and agent checks
Apache-2.0 judge on Granite 4.1 8B with a hybrid reasoning mode. Covers harm, groundedness and function-calling hallucination, plus custom criteria.
IBM Research released Granite Guardian 4.1 8B in April 2026; the Hugging Face repository was created on 16 April. It is a safety model that judges whether prompts or responses meet specified criteria, with a hybrid mode that can reason before scoring or return a direct yes/no. It is fine-tuned from Granite 4.1 8B, has 8B parameters and is licensed Apache-2.0.
Risk categories on the card are general (harm, social bias, jailbreaking, violence, profanity, unethical behaviour), RAG hallucination (context relevance, groundedness, answer relevance), agentic (function-calling hallucination) and custom bring-your-own criteria.
Reported results, non-think and think: out-of-distribution safety aggregate F1 0.79 and 0.78; LM-AggreFact groundedness balanced accuracy 0.760 and 0.764; FC Reward Bench 0.79 and 0.78; IFEval multi-constraint 0.844 and 0.827. Best-of-N reward modelling on JETTS scores 70.29, non-think only.
The limits are stated. It was trained and tested on English only. It must be used in its prescribed yes/no scoring format, and deviations can produce unexpected output. Reasoning traces may contain unsafe content and may not faithfully reflect its decision. Custom criteria need separate testing. For stricter cost or latency limits, IBM points to its 38M-parameter hate, abuse and profanity model.
A judge that also checks agent calls, which is a useful reach.
Sources
Editor's first pass (AI-assisted)
- Real 4.5: Open Apache-2.0 weights and a GGUF repo.
- Useful 4: One model for safety, RAG groundedness and agent function-call checks, with bring-your-own criteria.
- Notable 3.5: A long-running enterprise guard family continuing.
- Novel 3: Iterates on the earlier Granite Guardian line.
Comments