Mistral releases Shieldstral, a 3B Apache-2.0 multimodal safety classifier that takes its policy at inference time

Sources: announcement, weights, discussion
Summary: Mistral released Shieldstral 1.0, a 3B multimodal safety classifier under Apache-2.0, with weights on Hugging Face. The moderation policy is supplied at inference time rather than baked into the weights. Mistral states the model runs on a single 16GB GPU. Mistral also claims the model matches or outperforms open guard models up to seven times its size and sets a new state of the art on multimodal moderation. Those are Mistral's own figures, which no third party has reproduced, so they are carried as vendor claims and not as results.
Why it matters: Re-targeting a guardrail to a new deployment context becomes a prompt change rather than a retraining job, and a classifier that fits one 16GB GPU can sit inline per request instead of behind a moderation API.