Shieldstral: a moderation filter of 3 billion parameters that reads the rule when judging

Mistral publishes in Apache 2.0 a text and image security classifier that accepts a policy written in natural language for inference, rather than a fixed list of categories.

Moderation guardrails are usually trained on a fixed set of risk categories: to change policy, you have to retrain. Shieldstral, published by Mistral under the Apache 2.0 license on Hugging Face, takes a different approach to the problem — it receives the moderation policy in plain text at the time of inference and treats it as a question to answer about the submitted content.

The model has 3 billion parameters, covers twelve languages, judges both text and images, and fits on a single 16 GB GPU — so on a machine that many teams already own, which allows moderation to be kept in-house rather than sending it to a third-party API. Mistral claims 84.9% average F1 on text security, tied with GPT-OSS-Safeguard-20B, seven times larger, and 83.8% on multimodal security, ahead of all the benchmarks it evaluated.

This is the company's third moderation model, after two services available only in APIs, and the first that it publishes in open weights.