Skip to content
SPECIFICTION
Get listed

Granite Guardian

ibm.com

LinuxAPISelf-hosted
Visit ibm.com
Spec score 7.4 #20 of 27 in AI Content Moderation Software

What it is

Granite Guardian adds safeguards for prompts, responses, and agent workflows around AI models. It detects hallucinations, harmful content, bias, jailbreaking attempts, and retrieval-augmented generation quality issues. RAG checks cover context relevance, response groundedness, and answer relevance. Function calls can be assessed for syntactic and semantic hallucinations, including invalid calls or fabricated information. Granite Guardian 4.1 offers thinking mode with reasoning traces and non-thinking mode with a judgment score; the 4.1 8B model accepts user-defined criteria such as formatting, length, and domain instructions. Models use the Apache 2.0 license for research and commercial use and can run in vLLM or listed providers including watsonx.ai, Ollama, and LM Studio. IBM says they are trained and tested only on English data and should be used strictly in the prescribed scoring mode.

Spec sheet

Purpose
Granite Guardian adds specialized safeguards that detect risks in prompts and responses from any AI model.Read 9 Oct 2026 · ibm.com
Deployment
IBM describes Granite Guardian as deployable for any AI-native workload and provides instructions for running it in a vLLM container.Read 9 Oct 2026 · ibm.com
API
The IBM model selector shows a local vLLM deployment accepting requests at an OpenAI-compatible chat completions endpoint.Read 9 Oct 2026 · ibm.com
Security
IBM says Granite language models use cryptographic signatures to verify model provenance and undergo single-shot and adaptive multi-turn red teaming.Read 9 Oct 2026 · ibm.com
License
The models are released under Apache 2.0 for research and commercial use.Read 9 Oct 2026 · ibm.com
Headquarters
Armonk, New York, United StatesRead 28 Sep 2026 · ibm.com
Founded
1911Read 28 Sep 2026 · ibm.com
Agent workflows
It can assess function calls for syntactic and semantic hallucinations, including whether calls are valid and whether information is fabricated.Read 9 Oct 2026 · ibm.com
Custom criteria
Granite Guardian 4.1 8B supports user-defined judging criteria, including formatting rules, length constraints, and domain-specific instructions.Read 9 Oct 2026 · ibm.com
Deployment options
IBM lists watsonx.ai, Azure AI Foundry, Ollama, Replicate, LM Studio, Docker Hub, Dell, and AWS as providers or options for running Granite models.Read 9 Oct 2026 · ibm.com
Governance
IBM says its Granite AI Management System is ISO 42001 certified.Read 9 Oct 2026 · ibm.com
Hosted option
The IBM model selector lists watsonx.ai, Ollama, and LM Studio as other places the model is available.Read 9 Oct 2026 · ibm.com
Intended use and limitation
IBM says the main Guardian models are intended for moderate cost, latency, and throughput use cases, while smaller models can suit stricter cost, latency, or throughput requirements.Read 9 Oct 2026 · ibm.com
Local platforms
IBM's partner page describes LM Studio as running models locally on Mac, Windows, and Linux, and Ollama as supporting devices including CPUs, GPUs, Apple M-series MacBooks, and Raspberry Pi.Read 9 Oct 2026 · ibm.com
Model limitation
IBM documentation says Granite Guardian is trained and tested only on English data.Read 9 Oct 2026 · ibm.com
Model security
IBM says Granite language models have cryptographic signatures to verify model provenance and undergo single-shot and adaptive multi-turn red teaming using ARES.Read 9 Oct 2026 · ibm.com
Operating modes
Granite Guardian 4.1 supports a thinking mode with reasoning traces and a non-thinking mode that outputs a judgment score.Read 9 Oct 2026 · ibm.com
RAG checks
For RAG use cases, it evaluates context relevance, response groundedness, and answer relevance.Read 9 Oct 2026 · ibm.com
RAG evaluation
For RAG use cases, Guardian assesses context relevance, response groundedness, and answer relevance.Read 9 Oct 2026 · ibm.com
Risk detection
It detects hallucinations, harmful content, bias, jailbreaking attempts, and RAG quality and accuracy issues across prompts, responses, and agentic workflows.Read 9 Oct 2026 · ibm.com
Usage caution
IBM says Guardian models should be used strictly in the prescribed scoring mode and that reasoning traces may contain unsafe content and may not be faithful.Read 9 Oct 2026 · ibm.com
Use limitation
IBM says the models should be used strictly for the prescribed scoring mode, and that deviation may lead to unexpected, potentially unsafe, or harmful outputs.Read 9 Oct 2026 · ibm.com

Strong points and limits

Strong points

  • Detects hallucinations, bias, and harmful content
  • Evaluates RAG context and groundedness
  • Supports custom judging criteria
  • Apache 2.0 research and commercial license
  • Available through local and hosted options

Limits

  • English-only training and testing
  • Strict prescribed scoring mode required
  • Reasoning traces may contain unsafe content
  • Paid pricing model despite open-source note

Who it suits

It suits teams adding moderation, policy checks, or RAG and agent evaluations to AI applications. It also fits users needing local, self-hosted, Linux, or API deployment options.

Verdict

Granite Guardian focuses on configurable risk scoring for AI text, retrieval, and agent behavior. Deployment choices range from local tools to hosted providers, but IBM documents language and usage constraints.

Where Specifiction places it

Alternatives to Granite GuardianAll alternatives →