Mistral AI released Shieldstral on August 4, its first open-weights safety model: a 3-billion-parameter multimodal classifier that screens text and images against moderation policies written in plain language at inference time, rather than a fixed taxonomy baked into the weights. That design means one checkpoint can guard a cybersecurity research tool and a mental-health platform under different rules without retraining — the operator writes a yes/no question, an evaluation context and the content, and the model returns a calibrated score from a single token. Built on Ministral-3-3B with a Pixtral vision encoder and trained on about 54 million samples, Shieldstral scored an average F1 of 84.9% on text safety benchmarks — first among models from 4B to 20B parameters — and 83.8% on multimodal safety, ahead of the next-best OmniGuard-7B. It runs on a single 16GB GPU, covers 12 languages, and ships under Apache 2.0 on Hugging Face. Mistral positions the release as a critique of how guardrails are usually built and debuted it as an inaugural member of the Open Secure AI Alliance alongside NVIDIA, training the model on its Forge platform.
Mistral's Shieldstral Puts AI Safety Policies in the Prompt, Not the Weights
The Paris lab's 3B open-weights classifier judges text and images against plain-language policies at inference time, matching guard models up to seven times its size.
AI synthesis3 sources3
