Mistral's Shieldstral: 3B open-weights model for multimodal moderation

Mistral unveils Shieldstral, a 3B open-weights multimodal safety classifier that rivals models up to seven times its size. This model transforms content moderation into a policy-adaptive question-answering task, allowing plain-language policies to be applied at inference time. Unlike traditional guardrail models, Shieldstral accepts policies in natural language, unifying text and image safety evaluation without requiring retraining.

Available under the Apache 2.0 license, Shieldstral delivers calibrated safety scores across various benchmarks while running efficiently on a single 16GB NVIDIA GPU. The model's adaptability enables it to handle diverse content types, including text, images, and prompt-response pairs, with a single interface. Shieldstral's performance surpasses open guard models of similar size in text safety, refusal detection, policy adaptability, and multimodal benchmarks. trained on a diverse set of real and synthetic data, the model learns to distinguish specific policies from similar, easily confused ones, making it highly effective at adapting to novel policies at deployment time.