French AI company Mistral has open-sourced Shieldstral, a small model designed for content moderation. In Mistral’s official announcement, the company described Shieldstral as a 3 billion-parameter open-source multimodal safety classifier released under the Apache 2.0 license. It can run on a single 16GB NVIDIA GPU.
Moderation framed as a binary policy question
Mistral said Shieldstral takes a different approach from traditional content protection models. Those systems often rely on fixed built-in classification standards. Shieldstral instead turns moderation into a binary question-and-answer task: a user writes the policy in natural language, and the model evaluates whether a given piece of content violates that policy.
The content can be text, an image, or a combination of both. Shieldstral then returns a calibrated yes-or-no probability, and the process is completed in a single forward pass. One of the main benefits Mistral highlighted is that policies can be swapped or redefined at inference time, so changing the rules does not require retraining the model.
Mistral says the 3B model competes with much larger peers
On performance, Mistral said Shieldstral can match or even exceed open-source guard models about seven times larger across benchmarks for text safety, refusal detection, policy adaptation, and multimodal evaluation.
The model’s weights have been uploaded to Hugging Face. Mistral also said it is a founding member of the Open Secure AI Alliance, a group launched by organizations including NVIDIA. The alliance had previously included NVIDIA’s open-source agent framework as well.
Option for teams that want local moderation tools
For teams that want to build their own content moderation systems without sending data to the cloud, Shieldstral is positioned as a smaller and practical option.