MANY MINDEDMistral gave away the safety layer
Mistral has published Shieldstral-1.0-3B, a three-billion-parameter multimodal safety classifier, on Hugging Face under the Apache 2.0 licence. It evaluates text, images, or both, as a binary question-answering task: an instruction, a yes-or-no question about whether content is acceptable, and the document, returning a calibrated probability. It runs on a single 16GB GPU.
The design choice that matters is that it is policy-adaptive. It takes plain-language policies at inference time, so the same checkpoint serves different deployments without retraining. Content moderation systems have historically needed a new fine-tune for every new rule, which is why the capability has been a paid service rather than a component.
The abundance argument here is unusually clean. Content classification is not glamorous, but it is a mandatory tax on anyone operating a platform, a forum, a game, a school system or a marketplace — and it has been priced per API call by the handful of companies that could train the models. Apache 2.0 weights that run on hardware costing a few hundred dollars take a recurring per-call cost to roughly zero for anyone willing to host it. That is the pattern this project keeps documenting: a capability moves from rented to owned, and the price of the rented version has to answer for it.
The claim that it matches or outperforms open guard models up to seven times its size is Mistral's own benchmark, across four self-defined categories, published as charts rather than numbers. Nobody outside the company has replicated it. The licence and the hardware floor are facts; the ranking is a vendor's account.
Source: Mistral AI