The model's weights are available under Apache 2.0 for developers to inspect and deploy.
What happened
Shieldstral treats moderation as a question about text, images or combined content. Instead of choosing only from fixed categories, developers can supply a natural-language policy and receive a safety score without retraining the model.
Why it matters
Mistral says the model runs on a single 16GB GPU and reports strong results against larger guard models. Those comparisons describe its selected benchmarks, not a guarantee for every product or policy.
The bigger picture
Our take. Adaptable policies can make moderation easier to maintain. Teams still need representative examples, calibrated thresholds and a way to review mistakes before relying on a classifier in a real service.
Source checked November 18, 2026. Company claims are attributed; paragraphs marked “Our take” are editorial interpretation.