Shieldstral
Mistral's small open model that checks text and images against a safety rule you write yourself in plain language.
Shieldstral is a 3 billion parameter model that Mistral released in August 2026 with open weights under an Apache 2.0 licence. It does not chat. Its only job is to read a piece of content and answer one yes or no question about it, such as “does this promote violence”, and return a confidence score. Because you write that question yourself, the same model can enforce whatever rules your product needs without being retrained.
It reads images as well as text, covers twelve languages, and is small enough to run on a single 16 GB graphics card, which is why it is aimed at companies that want their moderation to happen on their own machines rather than through someone else’s API.