Shieldstral 1.0

Shieldstral 1.0 is Mistral AI's 3B open-weight multimodal safety classifier, released August 4, 2026 under Apache 2.0. Built on Ministral-3-3B-Base-2512 with a Pixtral vision encoder, it moderates text, images and text-image combinations.

Its signature design is policy adaptivity: instead of fixed moderation categories baked into weights, the developer passes a plain-language policy as input and the model returns a continuous safety score from a single forward pass. Retargeting to a new policy needs no retraining.

Mistral reports it matches or beats open guard models up to 7x its size on text safety and multimodal benchmarks, and it runs on a single 16GB GPU. Weights are on Hugging Face at mistralai/Shieldstral-1.0-3B.