Model reference · open weights

shieldgemma-2

shieldgemma-2 is an open-weight language model from Google. shieldgemma-2-4b-it (BF16) weighs 8.6 GB; the smallest configuration that runs it is RTX 3060 12 GB.

• ShieldGemma 2 is a 4.3 billion parameter image-text-to-text model developed by Google for image safety classification.
• It is designed to function as an input or output filter for vision language models and image generation systems, checking content against policies for sexually explicit, dangerous, and violent material.
• The model is trained on the Gemma 3 4B IT checkpoint and is released under the gemma licence.

Summary of the google/shieldgemma-2-4b-it model card, 2026-10-01

Hugging Face ↗GitHub ↗Paper ↗