# ShieldGemma-9B ![Organization](https://img.shields.io/badge/Org-Google-555) ![Pricing](https://img.shields.io/badge/Price-Open%20Weights-2da44e) ![License](https://img.shields.io/badge/License-Gemma%20Terms-d29922) ![Modality](https://img.shields.io/badge/Modality-Text-8957e5) > **Quick answer:** ShieldGemma-9B is Google's 9B open-weights text safety classifier, built on the Gemma architecture and released alongside the original Gemma model family. It trails the newer, smaller Shieldstral, Qwen3Guard, and GPT-OSS-Safeguard models across most benchmarks in Mistral's comparison, reflecting its earlier (2024) release relative to those 2025-2026 classifiers. ## At a Glance **Where it leads:** Not the top performer on any benchmark in Shieldstral's comparison set. **Where it lags:** Notably behind on WildGuardTest (Prompt) and HarmBench (Prompt), where newer guard models have made substantial gains. **Best for:** Teams already using Gemma-family models who want a same-family, open-weights baseline classifier. ## What ShieldGemma-9B Is ShieldGemma-9B is one of Google's earliest dedicated open-weights safety classifiers, built on the Gemma architecture to classify prompts and responses against a defined harm taxonomy (including categories like sexually explicit content, hate speech, harassment, and dangerous content). ## Specifications | Field | Value | |---|---| | Organization | Google | | Parameters | 9B | | Base model | Gemma | | License | Gemma Terms of Use | | Modality | Text only | ## Pricing Open weights, free to download under Gemma's Terms of Use. ## Public Benchmark Scores *Scores from Mistral AI's Shieldstral model card, shown for context. Not Benchgen measurements.* | Benchmark | Score | |---|---| | [WildGuardTest (Prompt)](/benchmarks/allen-institute-for-ai/wildguardtest-prompt) | 46.0% | | [HarmBench (Prompt)](/benchmarks/center-for-ai-safety/harmbench-prompt) | 50.2% | ## Related Models & Benchmarks - [ShieldGemma-2-4B](/models/google/shieldgemma-2-4b) · [Shieldstral 1.0](/models/mistral-ai/shieldstral-1-0) - [WildGuardTest (Prompt)](/benchmarks/allen-institute-for-ai/wildguardtest-prompt) · [HarmBench (Prompt)](/benchmarks/center-for-ai-safety/harmbench-prompt) ## Frequently Asked Questions

What is ShieldGemma-9B?

Google's 9B open-weights text safety classifier built on the Gemma architecture.

Is it multimodal?

No, ShieldGemma-9B is text-only (see ShieldGemma-2-4B for Google's multimodal successor).

Is it open source?

Yes, released under Google's Gemma Terms of Use.

--- *Scores sourced from Mistral AI's [Shieldstral model card](https://huggingface.co/mistralai/Shieldstral-1.0-3B), shown for context. Last updated 2026-08-12.*

ShieldGemma-9B — Evaluations

Evaluations