- Home
- AI Content Moderation Software
- Granite Guardian
Granite Guardian
Armonk, New York, United States
What it is
Granite Guardian adds safeguards for prompts, responses, and agentic workflows across AI models. It detects hallucinations, harmful content, bias, jailbreaking attempts, and retrieval-augmented generation quality or accuracy issues. For RAG applications, it evaluates context relevance, response groundedness, and answer relevance. It can assess function calls for syntactic and semantic hallucinations, including invalid calls or fabricated information. Granite Guardian 4.1 8B supports user-defined judging criteria such as formatting, length, and domain-specific instructions. Granite Guardian 4.1 offers a thinking mode with reasoning traces and a non-thinking mode that returns a judgment score. The Apache 2.0 models can run in a vLLM container or through listed options including watsonx.ai, Ollama, LM Studio, Azure AI Foundry, Replicate, Docker Hub, Dell, and AWS. IBM says they are trained and tested only on English data.
For and against
For
- Detects harmful content, bias, and jailbreaking
- Evaluates RAG relevance and groundedness
- Supports user-defined judging criteria
- Offers thinking and non-thinking modes
- Apache 2.0 license permits research and commercial use
Against
- Trained and tested only on English data
- Models must be used in prescribed scoring mode
- Reasoning traces may contain unsafe content
- Pricing model is paid despite open-source note
The verdict
Granite Guardian focuses on configurable safety and quality judgments around AI interactions. Deployment spans local tools, containers, and hosted providers, but IBM’s usage restrictions and English-only training are important constraints.
Who it suits
It suits organizations that need moderation, custom policy checks, RAG evaluation, or agent-workflow safeguards. It can fit moderate cost, latency, and throughput use cases, with smaller models for tighter constraints.
On the record
- Agent workflows
- It can assess function calls for syntactic and semantic hallucinations, including whether calls are valid and whether information is fabricated.ibm.com · read 9 Oct 2026
- API
- The IBM model selector shows a local vLLM deployment accepting requests at an OpenAI-compatible chat completions endpoint.ibm.com · read 9 Oct 2026
- Custom criteria
- Granite Guardian 4.1 8B supports user-defined judging criteria, including formatting rules, length constraints, and domain-specific instructions.ibm.com · read 9 Oct 2026
- Deployment
- IBM describes Granite Guardian as deployable for any AI-native workload and provides instructions for running it in a vLLM container.ibm.com · read 9 Oct 2026
- 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.ibm.com · read 9 Oct 2026
- Governance
- IBM says its Granite AI Management System is ISO 42001 certified.ibm.com · read 9 Oct 2026
- Hosted option
- The IBM model selector lists watsonx.ai, Ollama, and LM Studio as other places the model is available.ibm.com · read 9 Oct 2026
- 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.ibm.com · read 9 Oct 2026
- License
- The models are released under Apache 2.0 for research and commercial use.ibm.com · read 9 Oct 2026
- 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.ibm.com · read 9 Oct 2026
- Model limitation
- IBM documentation says Granite Guardian is trained and tested only on English data.ibm.com · read 9 Oct 2026
- 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.ibm.com · read 9 Oct 2026
- Operating modes
- Granite Guardian 4.1 supports a thinking mode with reasoning traces and a non-thinking mode that outputs a judgment score.ibm.com · read 9 Oct 2026
- Purpose
- Granite Guardian adds specialized safeguards that detect risks in prompts and responses from any AI model.ibm.com · read 9 Oct 2026
- RAG checks
- For RAG use cases, it evaluates context relevance, response groundedness, and answer relevance.ibm.com · read 9 Oct 2026
- RAG evaluation
- For RAG use cases, Guardian assesses context relevance, response groundedness, and answer relevance.ibm.com · read 9 Oct 2026
- Risk detection
- It detects hallucinations, harmful content, bias, jailbreaking attempts, and RAG quality and accuracy issues across prompts, responses, and agentic workflows.ibm.com · read 9 Oct 2026
- Security
- IBM says Granite language models use cryptographic signatures to verify model provenance and undergo single-shot and adaptive multi-turn red teaming.ibm.com · read 9 Oct 2026
