Mistral releases Shieldstral for multimodal moderation
Shieldstral is a 3B open-weight safety classifier that lets teams write moderation policy in the prompt instead of retraining the model.
Mistral released it under Apache 2.0, with support for text, images, and mixed text-image inputs through one interface. It can run on a single 16GB NVIDIA GPU, and the weights are available on Hugging Face. The model frames moderation as a yes-or-no question, then turns the yes/no logits into a continuous safety score for thresholding or ranking. Mistral says it matches or beats much larger open guard models across text safety, refusal detection, policy adaptation, and multimodal moderation. TestingCatalog's note
Mistral released it under Apache 2.0, with support for text, images, and mixed text-image inputs through one interface. It can run on a single 16GB NVIDIA GPU, and the weights are available on Hugging Face. The model frames moderation as a yes-or-no question, then turns the yes/no logits into a continuous safety score for thresholding or ranking. Mistral says it matches or beats much larger open guard models across text safety, refusal detection, policy adaptation, and multimodal moderation. TestingCatalog's note
score 6