Mistral AI releases Shieldstral, an open-weight safety classifier

Mistral AI released Shieldstral on August 4, a 3-billion-parameter open-weight multimodal safety classifier the company says outperforms models up to seven times its size by treating content moderation as a policy-adaptive question-answering task. Released under an Apache 2.0 license, Shieldstral accepts plain-language moderation policies at inference time rather than requiring retraining, unifies text and image safety scoring, and is designed to run on a single 16GB GPU.