AI & Enterprise
Mistral launches 3 billion-parameter safety model ShieldStral
Mistral AI has unveiled ShieldStral, a lightweight multimodal safety model that classifies AI model inputs and outputs. The open-weights model lets developers enter policies in natural language at runtime and returns a yes-or-no decision with a safety score. It can process text and images without additional retraining. The company said it was trained to distinguish similar criteria and reported benchmark scores averaging 84.9 percent for text and 83.8 percent for images.
