Mistral Releases Shieldstral 1.0 3B Safety Classifier for Flexible Content Moderation

Mistral AI has released Shieldstral 1.0 3B, an open-weights multimodal safety classifier designed for flexible content moderation. Unlike traditional safety systems that embed fixed harm categories into weights, Shieldstral accepts policies as plain-language questions at inference time and returns calibrated safety scores. The 3B model, built on Ministral-3-3B-Base with a native Pixtral vision encoder, achieves 84.9% average F1 score on text safety, matching larger models like GPT-OSS-Safeguard-20B, and 83.8% on multimodal safety. The model runs locally on a single GPU with 16GB VRAM in BF16 precision and is licensed under Apache 2.0 for commercial use. Training leveraged 54.1M samples including open-source text, synthetic contrastive data generated by rewriting safe text into targeted unsafe variants, and multimodal content. Serving infrastructure is already integrated with vLLM, llama.cpp, SGLang, and Transformers.