Mistral AI, a French company, has introduced a safety model named Shieldstral designed to simplify the moderation of AI-generated content.
Shieldstral is a multimodal model that accepts text and/or images as input, although it does not support audio or video input.
The model operates using a flexible question-answering approach where operators formulate moderation guidelines as natural language Yes/No questions. It provides a safety score based on the probabilities of 'Yes' and 'No' logits, which are normalized via Softmax into a continuous safety value. Shieldstral does not provide reasoning or justification, only a score.
Performance and technical specifications
Shieldstral-1.0-3B is built on the Ministral-3B-Base-2512 causal language model and utilizes a Pixtral-Vision-Encoder for image processing. The model has approximately 4 billion parameters, specifically 3.8 billion parameters.
Shieldstral obtained an F1 score of 84.9, matching the 20B GPT-OSS-Safeguard model.
Mistral AI claims Shieldstral achieves the same performance as models seven times its size. The company states that Shieldstral is more performant than any competitor for a lower weight in text and multimodal security tests, though it is surpassed by others in refusal detection and policy adaptability.
Usage and deployment
The model can process only one moderation policy per call; multiple policies require separate requests. While its theoretical context window is 256k, Mistral recommends limiting it to 32k.
For local operation in bfloat16 format, Shieldstral requires a minimum of 8 GB of memory, though Mistral recommends 16 GB. It can run on a 16 GB Nvidia GPU.
Shieldstral is available as an open weights model under the Apache-2.0 license on Hugging Face. The Hugging Face model card describes Shieldstral as a « public preview ».
