Mistral ships a safety model that takes plain-English rules

Mistral released Shieldstral on August 4, a 3B-parameter safety classifier that scores content against rules written in plain language instead of fixed categories.

That's useful for anyone whose moderation policy changes by jurisdiction or product surface. A new rule is just a rewritten sentence, not a retrained classifier.

Policy evaluation happens at inference time: the model reads the rule and returns a calibrated score from a single token, no fine-tuning required.

The company claims Shieldstral matches open guard models up to seven times its size on text, per the announcement. It also sets a new standard on multimodal moderation across 12 languages.

It runs on one 16GB GPU and handles text and images through the same interface.

Apache 2.0 licensing means teams can inspect, modify, or self-host the model instead of calling a hosted moderation API.

Shieldstral also marks the company's first contribution to the Open Secure AI Alliance, the Nvidia-led group that launched July 28 with 52 partners.