Mistral's new open 3B Shieldstral model evaluates AI inputs and outputs through natural-language yes-or-no safety questions rather than a fixed taxonomy. According to The Decoder, operators can define criteria at runtime, retain control over policy categories, and run the model locally. The report says Shieldstral matches safety models roughly seven times larger on some benchmarks. However, the supplied material does not identify those benchmarks, provide scores, name the comparison models, or specify licensing and deployment requirements, so the performance claim still needs verification against Mistral's primary documentation.
No heat snapshots are available in the last 24 hours.