Granite Guardian
Granite Guardian adds safeguards that identify risks in prompts and responses from AI models.
What Granite Guardian is
Granite Guardian adds safeguards that identify risks in prompts and responses from AI models. It can flag hallucinations, harmful content, bias, jailbreaking attempts, and RAG quality or accuracy issues, including within agentic workflows. For RAG, it evaluates context relevance, response groundedness, and answer relevance. It can also assess function calls for invalid calls or fabricated information. Granite Guardian 4.1 offers a thinking mode with reasoning traces and a non-thinking mode that returns a judgment score. The 4.1 8B model supports user-defined judging criteria, such as formatting rules, length limits, and domain-specific instructions. The models are released under the Apache 2.0 license for research and commercial use, and can be deployed in a vLLM container or through listed options including watsonx.ai, Ollama, and LM Studio. IBM says the models are trained and tested only on English data and should be used strictly in their prescribed scoring mode. Reasoning traces may contain unsafe content and may not be faithful.
Good
- Checks for hallucinations, harmful content, bias, and jailbreak attempts.
- Evaluates RAG context, groundedness, and answer relevance.
- 4.1 8B supports user-defined judging criteria.
- Apache 2.0 license covers research and commercial use.
Watch out
- Models are trained and tested only on English data.
- IBM advises using models only in prescribed scoring mode.
- Reasoning traces may contain unsafe or unfaithful content.
Who it suits
Granite Guardian suits teams adding risk checks to AI prompts, responses, RAG systems, or agent workflows. It also fits users who need custom judging criteria or self-hosted deployment options.
Verdict
Granite Guardian offers configurable checks across prompts, responses, RAG, and agent workflows. Its English-only training and prescribed-mode cautions are important constraints to consider.
The facts
Each one read from Granite Guardian's own pages, with the date it was checked.
What it does
- Agent workflows
- It can assess function calls for syntactic and semantic hallucinations, including whether calls are valid and whether information is fabricated.checked 9 Oct 2001
- API
- The IBM model selector shows a local vLLM deployment accepting requests at an OpenAI-compatible chat completions endpoint.checked 9 Oct 2001
- Custom criteria
- Granite Guardian 4.1 8B supports user-defined judging criteria, including formatting rules, length constraints, and domain-specific instructions.checked 9 Oct 2001
- Deployment
- IBM describes Granite Guardian as deployable for any AI-native workload and provides instructions for running it in a vLLM container.checked 9 Oct 2001
- 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.checked 9 Oct 2001
- Governance
- IBM says its Granite AI Management System is ISO 42001 certified.checked 9 Oct 2001
- Hosted option
- The IBM model selector lists watsonx.ai, Ollama, and LM Studio as other places the model is available.checked 9 Oct 2001
- 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.checked 9 Oct 2001
- License
- The models are released under Apache 2.0 for research and commercial use.checked 9 Oct 2001
- Model limitation
- IBM documentation says Granite Guardian is trained and tested only on English data.checked 9 Oct 2001
- 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.checked 9 Oct 2001
- Operating modes
- Granite Guardian 4.1 supports a thinking mode with reasoning traces and a non-thinking mode that outputs a judgment score.checked 9 Oct 2001
- Purpose
- Granite Guardian adds specialized safeguards that detect risks in prompts and responses from any AI model.checked 9 Oct 2001
- RAG checks
- For RAG use cases, it evaluates context relevance, response groundedness, and answer relevance.checked 9 Oct 2001
- RAG evaluation
- For RAG use cases, Guardian assesses context relevance, response groundedness, and answer relevance.checked 9 Oct 2001
- Risk detection
- It detects hallucinations, harmful content, bias, jailbreaking attempts, and RAG quality and accuracy issues across prompts, responses, and agentic workflows.checked 9 Oct 2001
- Security
- IBM says Granite language models use cryptographic signatures to verify model provenance and undergo single-shot and adaptive multi-turn red teaming.checked 9 Oct 2001
- 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.checked 9 Oct 2001
- 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.checked 9 Oct 2001
The company
- Founded
- 1911checked 28 Sep 2001
- Headquarters
- Armonk, New York, United Stateschecked 28 Sep 2001
Where it runs
- 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.checked 9 Oct 2001
Alternatives to Granite Guardian
- 1
Amazon Rekognition Content Moderation
Free planWebAPI7.5From $0.10 a month - 2
WebPurify
Free planAndroidiOSWebAPI7.4From $5 a month - 3
Tisane API
Free planWebAPILinuxmacOS7.1From $99 a month - 4
Alibaba Cloud Content Moderation
iOSWebAPI6.9From $0.09 a month - 5
Hive Moderation
Free planWebAPIExtension6.8Free plan - 6
Sightengine
Free planWebAPI6.8From $29 a month