Shieldstral

Shieldstral is Mistral's content-safety classifier: a 3-billion-parameter multimodal model that takes a piece of text or an image, plus a moderation policy written in ordinary English, and returns a verdict on whether the content violates that policy. It shipped on August 4, 2026 as Apache 2.0 open weights with an accompanying technical report, and Mistral puts the hardware floor at a single 16GB GPU. There is no hosted API SKU at launch — you run it yourself or you don't run it.

The design choice that separates it from the moderation endpoints most teams currently call is where the policy lives. OpenAI's moderation API and similar services score against a category taxonomy fixed in the weights; changing what counts as a violation means filing a request with a vendor or fine-tuning your own model. Shieldstral reads the rule at inference time, so a children's education app and a security research forum can load identical weights and enforce opposite definitions of "instructions for making something dangerous." Changing enforcement becomes editing a string. The flip side is that the policy text rides along in every request, so prompt length grows with your rule count, and small rewordings can shift classifier behaviour in ways nobody predicts from reading the diff — your policy file now needs the same review discipline as the code around it.

The headline benchmark claim is that Shieldstral matches or beats models up to seven times its size across text safety, refusal detection, policy adaptability and multimodal tests. Classification is exactly the workload where small specialised models close the gap on general-purpose giants, so the shape is plausible, but all four numbers were measured by Mistral on a task mix Mistral assembled, and "policy adaptability" is a dimension introduced in the same document as the model built to win it. Treat that one as marketing until someone outside the lab reproduces it. Worth it if you're already running your own inference and want user content to stay inside your process; skip it if you handle a few hundred messages a day, where renting a per-call moderation API beats paying rent on an idle 16GB card.