Shieldstral 1.0 is a compact multimodal safety model from Mistral AI. It judges text, images or a combination of both against a policy supplied in natural language and returns a continuous safety score. Because the policy is given at inference time, the model can be pointed at new rules without retraining. It is intended for lightweight, real-time moderation of user prompts, model responses and refusals, including on edge devices. The model card lists support for 12 languages. The model is delivered as open weights and runs with tools such as vLLM, llama.cpp, SGLang or Hugging Face Transformers.
Mistral AI is a French company and therefore subject to European law. The weights are on Hugging Face under the Apache 2.0 licence, for commercial and non-commercial use. Because you can run the model in your own environment, prompts and responses need not reach Mistral. Mistral AI remains the party that maintains and releases the model. The model card cautions that reliability varies across languages and domains and that adversarial or obfuscated inputs can reduce accuracy.