Mistral releases model with new technique for safety classifications

Mistral AI has presented Shieldstral, a safety model designed to simplify the moderation of AI-generated content. It uses a flexible question-answer approach instead of fixed categories: operators formulate their moderation policies as natural-language yes/no questions, such as "Does this content call for real violence?", and Shieldstral returns a safety score based on which the prompt or AI response can be approved or rejected. According to Mistral, this method has two advantages: first, operators can easily adjust the questions asked about the content at any time; second, it can better assess the context of content. For example, what might be necessary original quotes from contemporary speeches in a school lecture about concentration camps would be highly critical as content for a private blog post. The authors of the associated paper argue that previous guardrail models with fixed taxonomies cannot adequately capture such contextual differences because permissibility is often tied to categories. Input treats moderation as a binary question-answering task: the model receives a system instruction, context, and a yes/no question, and during inference outputs only the yes and no logits, normalized via softmax into a continuous safety score. Mistral claims Shieldstral achieves the same performance as seven times larger models but runs on a 16 GB Nvidia GPU. Technically, Shieldstral is based on the Ministral-3B-Base-2512, a causal language model from Mistral's own family, which was recently expanded with Leanstral 1.5, optimized for mathematical proofs and formal verification. For image processing, a Pixtral-Vision-Encoder is used. The 3-billion-parameter model is available for download as Open Weights under the Apache-2.0 license on Hugging Face.