Mistral AI Unveils Shieldstral 1.0 3B: Open-Weights Multimodal Safety Classifier Outperforming Models 7x Its Size

Mistral AI has released Shieldstral 1.0 3B, an open-weights, policy-adaptive multimodal safety classifier that treats content moderation as a single yes/no question rather than a fixed taxonomy of harm categories. Unlike most guardrail models that bake harm categories into their weights, Shieldstral lets operators write the policy as a plain-language question at inference time, returning a calibrated safety score from a single forward pass. Built on Ministral-3-3B-Base-2512 with a native Pixtral vision encoder and released under Apache 2.0, it reports 84.9% average F1 on text safety, matching GPT-OSS-Safeguard-20B, and 83.8% on multimodal safety, ahead of all baselines Mistral evaluated. The model fits in 16GB of VRAM in BF16, runs on a single GPU, and supports vLLM, llama.cpp, SGLang, and Transformers. Its training data includes roughly 54.1M samples, with contrastive generation creating hard negatives to teach policy-specific violations. Weaknesses include lower performance on low-resource languages like Arabic and Indonesian, and reduced reliability on adversarial or obfuscated inputs and very long documents.