Granite Guardian review
A focused open-source toolkit for screening LLM risks and evaluating AI outputs.
Reviewed by iTechGuides Editors · Editorial team · Updated Oct 2026
Granite Guardian is IBM’s family of open-source risk-detection models and LoRA adapters for teams evaluating AI prompts, responses, conversations, retrieval-augmented generation, and agentic function calls. It is designed for LLM teams that need screening for harmful content, bias, profanity, violence, sexual-content risks, jailbreak attempts, hallucinations, and groundedness issues. The models are released under the Apache 2.0 license and support API and self-hosted deployment options.
According to IBM’s published pricing page, the Granite Guardian 3.8B plan costs $0.212 per 1M tokens for pay-as-you-go model inference through IBM watsonx.ai. The ecosystem also includes Ollama, vLLM, and Hugging Face, giving teams several runtime options alongside watsonx.ai. Its Bring Your Own Criteria capability lets teams define custom evaluation rules, while Think and no-think modes and LoRA adapters support different evaluation and downstream-task requirements.
Granite Guardian’s main strength is depth within LLM risk evaluation rather than breadth across every moderation format. It covers prompt and response screening, jailbreak detection, RAG context relevance, groundedness, answer relevance, and agentic function-call hallucination detection in one focused model family. That makes it a strong fit for teams building or assessing text-based AI systems, especially where custom policies and local execution matter. Teams seeking broad multimodal moderation should consider a product with a wider media scope instead.
Granite Guardian pros and cons
- Where it wins
- Screens prompts and responses for harmful content, bias, jailbreaks, and hallucinations
- Evaluates RAG relevance, groundedness, answers, and agentic function calls
- Supports custom criteria, local deployment, and IBM watsonx.ai integration
- Where it doesn't
- Focused on text and LLM evaluation rather than general multimodal moderation
- Granite Guardian 3.8B inference is billed per 1M tokens through watsonx.ai
- Requires teams to choose and operate an API or self-hosted deployment path
Granite Guardian fact sheet, pricing and score →
Advertiser disclosure: iTechGuides is reader-supported. Vendors can pay for top positions in our rankings and for a place on other products' pages, and we may earn a commission when you click some links. Bought places stand at the number bought; this page and our disclosure say so. Paying vendors may get a fuller, more detailed review from our editors, and that review can change their score. How we rank.
Last updated · How we research and update