Mistral shipped a 3B open-weights moderation model that takes your policy as plain text
Shieldstral is a 3B Apache 2.0 multimodal safety classifier that reads your moderation policy at inference time and answers with a calibrated yes or no from a single forward pass, on one 16GB GPU.
One 16GB GPU, any policy you can write in a paragraph. Mistral just open sourced the moderation layer.
Mistral released Shieldstral on August 4, 2026: a 3 billion parameter, Apache 2.0, open weights safety classifier that takes your moderation policy as plain text at inference time. Instead of shipping fixed harm categories that need retraining when your rules change, you hand it the policy and the content, and it returns a calibrated yes or no probability from a single forward pass.
Moderation as a question, not a category list
Every guard model until now has baked its harm taxonomy into training. If your product needs a different line, say a kids' app versus a security research forum, you retrain or you tolerate the mismatch. Shieldstral frames moderation as binary question answering with three parts: an instruction context, a yes or no query, and the document to judge, which can be text, an image, or both. The same weights handle prompt classification, response moderation, refusal detection, and toxicity checks, because those are all just different questions.
The numbers Mistral is claiming
Mistral says the 3B model matches or outperforms open guard models up to 7 times its size and runs on a single 16GB NVIDIA GPU, which puts self-hosted moderation within reach of one mid-range card rather than a cluster. The weights are on Hugging Face under Apache 2.0, with a paper on arXiv. The honest caveat: the benchmark claims are Mistral's own, published alongside the release. They are plausible and the method is documented, but nobody independent has reproduced them yet.
The release is also a signal about where Mistral sits. It shipped as an inaugural member of the Open Secure AI Alliance alongside NVIDIA, and it extends the run of European open weights releases we have been tracking: the moderation layer, historically a closed API you rented from a US platform, is now an Apache-licensed download.
Why a build studio cares
Every AI product we ship needs a moderation answer, and the practical choice has been a paid moderation API with someone else's policy or nothing. A small model that takes the client's actual content policy as a prompt, runs on one GPU, and costs nothing to license changes that math, especially for products whose policies are jurisdiction-specific or unusual. It goes on the evaluation list this week.
Next step: read Mistral's announcement and the model card. If your product needs a moderation layer that follows your policy instead of a vendor's, write to us at hello@gattyworks.com.