Granite Guardian 3.1 8B

Granite Guardian 3.1 8B is IBM's open risk-detection model for prompts and responses, released in December 2024 in 8B and 2B sizes under Apache 2.0. It covers harm dimensions from the IBM AI Risk Atlas and, new in the 3.1 release, detects hallucinations in agentic workflows: it monitors function calls for syntactic and semantic hallucinations, checking whether an agent fabricated information or pulled the wrong numbers from a retrieved record.

The model is trained on human annotations plus synthetic data informed by internal red-teaming, and IBM reports it outperforms other open models in its class on standard safety benchmarks. It is available on Hugging Face, through watsonx.ai, and via Ollama.