tl;dr: Mistral released Shieldstral, a 3B Apache 2.0 multimodal safety classifier that accepts policies as plain-language yes/no questions at inference time, eliminating the need to retrain for new taxonomies. It outputs calibrated safety scores from yes/no token logits, handles text and images, and reportedly matches or beats guard models up to 7x its size. It runs on a single 16GB GPU and was trained by unifying heterogeneous datasets, using contrastive policy pairs to teach discrimination, and merging LoRA checkpoints via SLERP.