Shieldstral is Mistral AI's first open-weights safety classifier: a 3-billion-parameter model that decides whether a piece of text or an image breaks a rule the developer writes in plain language at inference time. Instead of a fixed list of harm categories fixed during training, moderation becomes a yes/no question — the model returns only the probabilities of "yes" and "no", which are normalised into a continuous safety score, so a changed policy needs a changed prompt, not a retraining run. Released on 4 August 2026 under Apache 2.0, it is built on the Ministral 3B base, covers twelve languages and judges text, images or both. Mistral reports it matches or beats open guard models up to seven times its size, including 88.1 F1 on WildGuardTest and 97.7 on the multimodal VLGuard set. It fits on a single 16 GB GPU, which puts a moderation layer within reach of teams that cannot route user content to someone else's API.