Release2026-08-04
Mistral released Shieldstral, a small content-moderation model that judges text and images against a safety policy written out in plain language at the moment of the check, rather than against fixed categories built into the model. The weights are freely downloadable under the permissive Apache 2.0 licence and the model runs on a single 16GB GPU. Using it requires hosting it, and Mistral's performance claims against larger open guard models are its own.
What changed
Guardrail models had fixed harm categories baked into their weights and needed retraining to fit a new context.
What it unlocks
Checking text and images against a policy written in plain language at the moment of the check, with no retraining.
- 3B parameters, Apache 2.0
- matches models up to 7x larger
- runs on a single 16GB GPU
What you need to act on it
- a 16GB GPU or hosted equivalent
- engineering work to self-host the weights
Sources