Mistral has released Shieldstral, a 3B parameter multimodal safety classifier released under the Apache 2.0 license. Unlike traditional guardrail models that rely on fixed taxonomies, Shieldstral treats content moderation as a policy-adaptive question-answering task. It processes requests using a tripartite structure consisting of an instruction (defining context and strictness), a query (a plain-language yes/no question), and the document (text, images, or multimodal pairs). During inference, the model performs a single forward pass, outputting calibrated yes/no logits that are softmax-normalized into a continuous safety score. This architecture allows users to implement custom policies at deployment time without retraining. To achieve high performance, the model was trained on a consolidated dataset of heterogeneous sources, using a specialized processor to unify varying label formats and strictness levels. The training methodology utilized contrastive pairs—engineered by LLMs to distinguish between subtly different policies—to prioritize reasoning over memorization. Through a combination of LoRA fine-tuning and SLERP merging of multiple checkpoints, the 3B model matches or outperforms guard models up to 7x its size. Optimized for efficiency, Shieldstral can run on a single 16GB NVIDIA GPU, providing a unified interface for text and image safety evaluation.