Mistral Launches Shieldstral, a 3B Open-Weights Multimodal Safety Classifier

Mistral AI has released Shieldstral, a 3B open-weights multimodal safety classifier that outperforms models up to 7x its size. Instead of baking a fixed taxonomy of harm categories into its weights, Shieldstral frames content moderation as a policy-adaptive question-answering task: you write the policy as a plain-language question at inference time, and the model returns a calibrated safety score. One interface covers text, image, and combined content, with no retraining required.

Each request has three parts: an Instruct block that sets the evaluation context and strictness, a Document block with the content to judge, and a simple binary verdict. At inference, the model reads out the yes and no logits and softmax-normalizes them into a continuous safety score. This unifies prompt classification, response moderation, refusal detection, and toxicity detection into a single problem, letting one checkpoint adapt to novel policies at deployment time.

Released under Apache 2.0 as an inaugural member of the Open Secure AI Alliance with NVIDIA, Shieldstral runs on a single 16GB GPU, matching or beating open guard models up to 7x its size across text safety, refusal detection, policy adaptability, and multimodal benchmarks. It returns calibrated probabilities rather than discrete labels, so teams can threshold by confidence. For founders shipping AI products, that means content safety that adapts to their audience without retraining.