Mistral released Shieldstral on August 4 as an open-weight classifier for text and image safety decisions. The announcement describes a three-billion-parameter model under Apache 2.0 that accepts a policy expressed as a question and returns a score for the corresponding yes-or-no judgment.
Policies become part of the request
Instead of assigning every input to one fixed collection of harm labels, the model evaluates content against the supplied policy and context. Mistral positions this as a way to adapt moderation to different applications without retraining for each taxonomy. It reports comparisons across text safety, refusal detection, policy adaptability, and multimodal evaluations.
The publisher describes the output as calibrated. Whether a particular score supports automatic blocking in a new product is still a question for that product’s data and consequences.
Version the policy with the model
Our analysis: policy wording becomes an operational input when a classifier interprets natural language. Changing the question can change the decision boundary even if the weights stay identical. Store the model version, policy text, threshold, and outcome together so disputed decisions can be investigated.
Build examples around nearby cases that should receive different decisions, including quotation, educational discussion, and images whose meaning depends on accompanying text. Evaluate false positives as well as missed violations; an excessively broad block can make a legitimate workflow unusable.
A deployment plan should identify which decisions can be automatic, which go to human review, and how reviewers correct mistakes. Run proposed policy changes against the same held-out cases before rollout, then monitor disagreement rates rather than treating a model score as a universal definition of safety.
- Introducing Shieldstral.
Mistral AI · Aug 4, 2026
See the original announcement for availability and release details.