Mistral AI has released Shieldstral 1.0 3B, an open-weights model that treats content moderation as a single yes/no question rather than relying on fixed harm categories. Built on Ministral-3-3B-Base-2512 with a Pixtral vision encoder, it allows operators to define policies via plain-language prompts at inference time without retraining.

  • Reports 84.9% average F1 on text safety, tying GPT-OSS-Safeguard-20B, and 83.8% on multimodal safety, leading evaluated baselines.
  • Trained on roughly 54.1M samples using sibling-contrastive rewrites to teach specific policy violations.
  • Fits in 16GB VRAM and emits a single continuous safety score via one forward pass.
  • Weaknesses include performance on low-resource languages, obfuscated inputs, and very long documents.

The model enables real-time, cost-effective moderation for diverse deployment contexts by allowing dynamic policy enforcement through prompt engineering.