Mistral AI dropped a "reviewing nuclear bomb" on August 4th - the Shieldstral content review model, with a total of 3B (3 billion) parameters, uses open weights and is released under the Apache 2.0 license, and has been launched on Hugging Face. It supports 12 languages, and the most attractive feature is that it can run on a single 16GB GPU, while Mistral claims its content safety performance can rival open models seven times larger in scale, and it has achieved SOTA (the current state-of-the-art) in the field of multimodal content review.

image.png

Most protection models have a common problem: they directly weld the harmful category system into the weights. When the product changes the usage scenario, developers often need to retrain. Shieldstral takes a different approach - writing the review policies into the input. The operator can write a "yes or no" question themselves, then add an assessment scenario and strictness explanation, after which the model will only output a calibrated safety score from a single output token.

image.png

In terms of specific mechanisms, it breaks down each review task into a binary question and answer. The input consists of three labeled fields: <Instruct> explains the assessment scenario and strictness, <Query> poses a "yes or no" question (such as "Does this content promote physical violence?"), and <Document> provides the content to be reviewed - it can be a prompt, an answer, a combination of both, or an image with optional text. During inference, the model only reads the logical values of the "yes" and "no" tokens, normalizes them into a continuous score using softmax, and then gives a binary judgment with 0.5 as the threshold. With this design, it can cover prompt classification, response review, refusal detection, and toxicity detection all at once, fitting multimodal review into the capacity of a consumer-grade GPU.

Model address: https://huggingface.co/mistralai/Shieldstral-1.0-3B