ShieldGemma

ShieldGemma is Google's family of open-weight content-safety classifiers, first released in mid-2024 on Gemma 2 bases in 2B, 9B, and 27B sizes, followed by ShieldGemma 2, a 4B multimodal guardian built on Gemma 3 that screens both text and images. The models output safe/unsafe verdicts with per-category confidence across policies like hate speech, harassment, dangerous content, and sexually explicit material.

Weights are open under the Gemma Terms of Use (commercial use allowed) and run in Transformers, Keras, JAX, and Ollama, making them a popular self-hosted alternative to closed moderation APIs. They support both input filtering (screening user prompts) and output filtering (checking generations), with tunable thresholds per deployment.