Shieldstral is a 3B-parameter policy-adaptive multimodal safety classifier designed to assess text and image-containing inputs against criteria supplied in natural language. Presented in an arXiv preprint, the model frames moderation as a binary yes-or-no question-answering task, seeking to replace rigid category taxonomies with a single adaptable safety score.
The central idea is significant for teams building moderation workflows across changing policies, products, and jurisdictions. Instead of requiring a separate fixed label for every type of prohibited or sensitive content, Shieldstral is designed to accept an operator's moderation criterion at inference time. The authors report that the system matches or exceeds much larger models on multimodal safety benchmarks, while also delivering strong text-safety results.
The model and its evaluation are detailed in the Shieldstral arXiv preprint, published July 28, 2026. The paper describes Shieldstral as being built on Ministral-3B, from Mistral AI's Ministral 3 family, positioning the work around a relatively compact model architecture rather than the largest available multimodal systems.
How Shieldstral approaches multimodal moderation
