Granite Guardian 3.2

Granite Guardian 3.2, released February 26, 2025, is IBM's family of open safety models in 3B, 5B, and 8B sizes, fine-tuned from Granite 3.2 instruct models on human annotations plus internal red-teaming data. Beyond standard prompt/response harm detection against the IBM AI Risk Atlas taxonomy, it covers two enterprise-specific jobs: RAG risk detection (context relevance, answer groundedness, answer relevance) and agentic function-calling risk detection (hallucinated tool calls and fabricated intermediate steps).

Like all Granite models it is Apache 2.0 — genuinely open-source licensed, not a custom community license — and IBM reported it outperforming other open models in its class on standard safety benchmarks. The 3B-A800M variant targets efficient deployment.

Honest caveats: benchmark claims are IBM's own, and the ecosystem (cookbooks, integrations) is smaller than Meta's Purple Llama. But for enterprises already standardizing on Granite, an Apache-2.0 guard model with RAG and agent-call coverage is a meaningful differentiator.