Mistral AI Releases Shieldstral, a 3B-Parameter Content Moderation Model

Mistral announces Shieldstral 1.0: a 3 billion parameter open weight content moderator. It runs on a regular GPU with 16 GB of memory and achieves 7x better filtering accuracy than models of similar size on both general (text/image) and specialized harmful content rejection tasks. The model weights are published under Apache 2.0 license on Hugging Face.

1Policy as a Text Prompt — No Fine-Tuning Required

What sets Shieldstral apart from other approaches is that the moderation policies are defined using simple text prompts, rather than being encoded in the model weights. In this way, the same checkpoint can be used for both text and image moderation, without needing to do any fine-tuning when switching contexts. This represents a major paradigm shift, as before, modifying the moderation policy would have required a time-consuming process of re-fine-tuning and engineering effort.

With Shieldstral, anyone with access to a 16 GB GPU server can build high quality content filtering systems, without having to rely on large compute resources. Moreover, since the model weights are open, it’s possible to deploy Shieldstral directly within your own product, without depending on any external APIs.