Introducing Shieldstral.

Shieldstral is a new, compact AI model designed for advanced content safety classification across both text and images. It uniquely treats moderation as a question-answering task, allowing users to define safety policies with plain-language questions at the moment of evaluation. This innovative method allows the model to outperform competitors up to seven times its size while running efficiently on a single GPU. As an open-weights model, it provides a calibrated safety score for any piece of content, adapting to new policies on the fly without needing to be retrained.