Shieldstral (by Mistral AI)

Shieldstral is an open-source, 3B-parameter multimodal safety classifier from Mistral AI that reimagines content moderation as a policy-adaptive question-answering task. Instead of hard-coded harm taxonomies, it accepts a plain-language policy question at inference time and returns a calibrated safety score—unifying text, image, and refusal detection in a single forward pass. Released under Apache 2.0 and backed by the Open Secure AI Alliance with NVIDIA, it matches or beats models up to 7x larger while running on a single 16GB GPU. The model’s key innovation lies in its training: heterogeneous datasets are converted into a unified instruction–query–document format, with contrastive policy pairs teaching discrimination rather than memorization, enabling generalization to novel policies without retraining.