Mistral has released Shieldstral, a 3B open-weight multimodal safety classifier for teams that need moderation rules tailored to a product, audience, or domain. Released under Apache 2.0, the model handles text, images, and combined text-image content through one interface and can run on a single 16GB NVIDIA GPU. Its weights are available from Mistral on Hugging Face.
Shieldstral turns moderation into a binary question-answering task. At inference time, a developer supplies the evaluation context, a plain-language yes-or-no policy question, and the content to assess, which can be a prompt, a response, a prompt-response pair, or an image with optional text. The model reads the yes and no logits and converts them into a continuous safety score, allowing applications to set thresholds or rank results by confidence. Policies remain in the prompt, so a single checkpoint can be retargeted without retraining.
Mistral says Shieldstral matches or exceeds open-guard models up to 7 times larger in text safety, refusal detection, policy adaptability, and multimodal moderation. The company also reports state-of-the-art multimodal results, with all evaluation samples held out from training. The same setup supports prompt classification, response moderation, refusal checks, and toxicity detection.
The model was trained on real and synthetic sources whose different labels and taxonomies were converted into a shared instruction-query-document format. Mistral used contrastive examples to teach distinctions between closely related policies, supplemented scarce visual safety data with general image datasets as negative examples, and filtered image-query pairs with a vision-language reranker. LoRA fine-tuning and SLERP then combined checkpoints focused on public-data calibration, fine-grained policy discrimination, and base-model instruction following.
Mistral built Shieldstral on Forge, its platform for training, aligning, and evaluating custom models. The release also advances the company's work as an inaugural member of the Open Secure AI Alliance alongside NVIDIA and other organizations. Mistral plans to extend Shieldstral with broader multilingual coverage, stronger support for long documents, and broader multimodal safety capabilities.