Mistral released Shieldstral, a small open-weights model that judges whether text or images breach a content rule. The rule is written as a plain-language yes/no question at the time of use, so the same model can be pointed at different policies without being retrained. It is published under the permissive Apache 2.0 licence and is small enough to run on one 16GB graphics card, but it must be self-hosted and wired in by engineers.
What changed
Most content-moderation models have a fixed list of harm categories built into them, so adapting to a new context required retraining.
What it unlocks
Checking text and images against a moderation rule written as a plain-language question, and changing that rule without retraining the model.
- 3 billion parameters
- matches models up to 7x its size
- runs on a single 16GB NVIDIA GPU
- Apache 2.0 licence
What you need to act on it
- a 16GB GPU or equivalent hosting
- engineering work to integrate the model
- mistral.ai2026-08-04