Mistral AI Releases Shieldstral 1.0: 3B-Parameter Open-Weights Multimodal Safety Classifier
On August 4, 2026, Mistral AI released Shieldstral 1.0, a 3B parameter open weights multimodal safety classifier for content moderation, as an inaugural member…
Published on MyPrivateClaw
Aug 4, 2026, 5:11 PM UTC
Coverage date
Aug 4, 2026
Last updated
Aug 4, 2026, 5:11 PM UTC
News summary
Shieldstral 1.0 launched on August 4, 2026 as a 3B parameter multimodal safety classifier developed by Mistral AI as part of the Open Secure AI Alliance with NVIDIA. The model architecture builds on Ministral 3 3B Base 2512 with a native Pixtral vision encoder, producing safety verdicts from a single forward pass. It supports text only, image only, and text+image moderation with up to 32k token context window within the trained range. Shieldstral formulates content moderation as a binary question answering task where safety policies are expressed in natural language at inference time, returning a continuous yes/no safety score without requiring retraining. This policy adaptive mechanism serves as a key differentiator. The model supports 12 languages: English, French, Spanish, German, Italian, Portuguese, Dutch, Chinese, Japanese, Korean, Arabic, and Russian. Multilingual coverage includ…