Introducing Shieldstral.

Mistral has announced the release of Shieldstral, a 3B open-weights, policy-adaptive multimodal safety classifier available under the Apache 2.0 license as an inaugural member of the Open Secure AI Alliance alongside NVIDIA and other organizations. According to Mistral, traditional guardrail models bake a fixed taxonomy of harm categories into their weights, requiring retraining to adapt to different deployment contexts. Shieldstral addresses this limitation by framing content moderation as a binary question-answering task where users write the policy as a plain-language question at inference time. The model evaluates an instruction, a yes/no query, and a document containing text, an image, or a prompt-response pair, then outputs a calibrated continuous safety score derived from softmax-normalized yes and no logits. Mistral states that Shieldstral matches or outperforms open guard models up to seven times its size across text safety, refusal detection, policy adaptability, and multimodal benchmarks while running on a single 16GB GPU. The model was trained using a combination of unified heterogeneous data, contrastive pairs to teach fine-grained policy discrimination, grounded image data, and checkpoint merging using LoRA and SLERP on Mistral's Forge platform. Shieldstral unifies prompt classification, response moderation, refusal detection, and toxicity detection into a single interface, offering developers an adaptable approach to AI safety moderation without requiring model retraining.