Mistral launches open multimodal safety model Shieldstral with adaptable moderation

Last week, French AI company Mistral introduced Shieldstral, a 3 billion parameter open weight multimodal safety classifier for content moderation, released under the Apache 2.0 license with model weights available on Hugging Face. Designed for adaptable moderation, Shieldstral lets developers define custom policies for different products, audiences, or domains instead of relying on fixed safety taxonomies.

The model processes text, images, and combined text and image inputs through a single interface. Moderation is framed as binary question answering, with developers supplying plain language policy questions and the content to review at inference time. Shieldstral converts yes and no logits into continuous safety scores, allowing applications to set their own thresholds or rank content by confidence without retraining or fine tuning. It supports prompt classification, response moderation, refusal detection, toxicity detection, and policy compliance checks.

Mistral says Shieldstral matches or outperforms larger open guardrail models in areas including text safety, policy adaptability, and multimodal moderation, while running on a single NVIDIA GPU with 16GB of memory. Training combined real and synthetic data with contrastive examples, vision language reranking, and checkpoint merging to improve policy discrimination and calibration. Mistral plans to expand the model with broader multilingual support, improved long document handling, and additional multimodal safety capabilities.