Shieldstral
by Mistral AI · Model · Language & code models
Mistral's 3B open-weights multimodal safety classifier; takes a plain-English yes/no policy at inference and returns a calibrated score in one forward pass
AccessOpen weightsNeeds~16 GBSize3B parametersLicenseApache-2.0
Evidence
- Aug 5Claimed to tie the top guardrail model on averaged text-safety F1 across a dozen-plus benchmarks and beat other guardrail models, despite being much smaller
- Aug 5Rewriting the policy to zero-tolerance flipped the same content from safe (score 0.085) to unsafe, without retraining; read intent, not just scary words