Mistral AI released Shieldstral, a content review model with 3B parameters, on August 4th. It uses open weights under the Apache 2.0 license and is available on Hugging Face. The model supports 12 languages and runs on a single 16GB GPU. Mistral claims its content safety performance rivals open models seven times larger and achieves SOTA in multimodal content review. Shieldstral writes review policies into the input instead of welding harmful category systems into weights. Operators can write a "yes or no" question, add an assessment scenario and strictness explanation, and the model outputs a calibrated safety score from a single output token. It breaks each review task into binary questions with three labeled input fields: <Instruct> for scenario and strictness, <Query> for the question, and <Document> for the content, which can be a prompt, answer, combination, or image with optional text. During inference, the model reads logical values of "yes" and "no" tokens, normalizes them into a continuous score using softmax, and gives a binary judgment with 0.5 as the threshold. This design covers prompt classification, response review, refusal detection, and toxicity detection, fitting multimodal review into consumer-grade GPU capacity.