A Natural-Language Policy for Moderating Text and Images

Shieldstral adapts moderation to each product’s policies, evaluates text and images, and runs on a single NVIDIA GPU with 16 GB of memory.

Shieldstral is a 3-billion-parameter multimodal safety classifier released with open weights under the Apache 2.0 license. It evaluates prompts, responses, complete exchanges, or images, with optional accompanying text.

Rather than imposing a fixed list of categories, the model receives the moderation policy as a natural-language question. A company can ask whether content promotes targeted violence, whether it is appropriate for minors, or whether an assistant correctly refused a request, without retraining the model each time the policy changes.

Each request combines the evaluation context, a binary question, and the content to be reviewed. Shieldstral then compares the probabilities associated with “yes” and “no” to produce a continuous score that can be adjusted according to the desired level of caution.

The same interface covers text safety, refusal detection, adaptation to new policies, and multimodal moderation. Mistral says the model can run on a single NVIDIA GPU with 16 GB of memory.

Training combines approximately 54.1 million real and synthetic examples drawn from different taxonomies and converted into a common format. Contrastive pairs teach the model to distinguish between closely related policies, while the visual data is expanded and filtered to reduce incorrect associations.

Early reactions reflect the familiar divide surrounding safety tools. Some developers mainly highlight the ability to keep moderation local, freely adapt the rules, and limit hardware requirements. Others view the product as a censorship tool, question Mistral’s priorities, or criticize the binary verdict for not providing a detailed explanation of why content was blocked.