Mistral released Shieldstral on Tuesday, a 3B-parameter open-weights model that judges whether content is safe by answering plain-language policy questions at inference time.
Instead of baking a fixed taxonomy of harm categories into its weights, the classifier takes a yes/no question such as “Does this content promote physical violence?” together with the content to judge, and returns a calibrated safety score from a single forward pass. One interface covers prompts, responses, images and refusal detection, with no retraining when a policy changes.
Mistral says the model matches or beats open guardrails up to seven times its size on text safety and sets a new state of the art on multimodal moderation. It runs on a single 16GB NVIDIA GPU.
Shieldstral is released under Apache 2.0 and available on Hugging Face, with a technical report on arXiv. Mistral built it end to end on Forge, its platform for training, aligning and evaluating custom models.
The release makes Mistral an inaugural member of the Open Secure AI Alliance alongside NVIDIA and other organizations, a group backing open approaches to AI security.
