Mistral released a free safety filter that you describe in plain English, and it fits on one graphics card
Shieldstral is a 3 billion parameter open-weights model that checks text and images against a policy you write yourself, in ordinary sentences, without any retraining. It runs on a single 16 GB GPU under an Apache 2.0 licence.
French lab Mistral has published Shieldstral, a small open model whose only job is to look at a piece of content and answer one yes or no question about it. You supply the question in ordinary English. Ask it “Does this content promote physical violence?” and it returns a score between 0 and 1. The weights are free to download under an Apache 2.0 licence, which allows commercial use, and the model fits on a single 16 GB graphics card.
That sounds mundane until you know how this normally works. Most safety filters, often called guardrail models, are trained on a fixed list of categories: hate speech, self harm, weapons, and so on. If your product needs a category the model was never taught, you either retrain it or you live without it. Shieldstral moves the policy out of the training and into the prompt. Mistral calls this policy adaptive: one checkpoint, any number of rules, changed whenever you like.
The model is built on Mistral’s own Ministral 3B base with the Pixtral vision encoder attached, so it reads images as well as text, and it covers twelve languages including German, French, Italian and Spanish. On the published benchmark tables it matches or beats open guard models several times its size, and on the multimodal safety test VLGuard it scores 97.7 on the F1 measure against 88.5 for the next best. A verdict is a single forward pass, meaning the model produces one token and stops, which is why it is cheap enough to run on every message in a live chat.
What is behind this. Every company that ships a chatbot suddenly needs a moderation layer, and the obvious options are an API from OpenAI or Google, or a large open model that eats a lot of hardware. A 3B model that runs on one card and follows written rules lands neatly between them, and it lets a European company keep the checking on its own machines. Mistral, which sells hosted models too, has a clear interest in being the lab that makes the compliance piece free.
What this means for you: if you only use AI tools, nothing changes today, though filters like this are the reason a chatbot sometimes declines a question. If you build or run anything with a chat box in it, a school platform, a support inbox, a community forum, this is a serious option that costs nothing in licence fees. Two honest caveats from Mistral’s own model card: reliability varies by language, and deliberately obfuscated text, such as unusual spellings or encoded characters, can slip past it. Treat it as a filter, not a guarantee.
Sources
An 80 billion parameter model now runs on a normal Mac in 4.3 GB of memory, and a 35B one runs on an iPhone
Swiftlet is an open runtime that streams a model's expert weights from the SSD instead of loading them into memory. A Qwen model with 80 billion parameters peaks at 4.3 GB of RAM on a Mac. The trade off is speed and a large download.
