Mistral has unveiled Shieldstral, a 3B open-weights multimodal safety classifier designed for content moderation. It frames moderation as a policy-adaptive question-answering task, allowing plain-language policies to be applied at inference time without retraining. This approach unifies text and image safety evaluation.

Shieldstral outperforms models up to 7x its size on text safety and sets a new standard for multimodal moderation. It efficiently runs on a single 16GB NVIDIA GPU and provides calibrated safety scores across various benchmarks. The model accepts natural language queries about content, returning a continuous safety score based on a yes/no probability.

Released under Apache 2.0, Shieldstral supports diverse content types, including text, images, and prompt-response pairs. Its adaptive nature allows it to target new policies at deployment without requiring model retraining. Mistral emphasizes its commitment to open AI security with this release, inviting community contributions.