According to StatSense monitoring, European AI unicorn Mistral has launched a content moderation model called Shieldstral. The model has only 3 billion parameters and can inspect text, images, and multimedia content, as well as determine whether AI should refuse to respond.
Many traditional moderation models predefine categories such as pornography, violence, and hate speech. After a change in moderation criteria, it usually requires additional data supplementation, model retraining, or fine-tuning.
Shieldstral can directly comprehend natural language rules. Developers can modify the moderation requirements to switch standards without having to retrain the model.
The model supports 12 languages, including Chinese, and is licensed under Apache 2.0. The company stated that a single 16GB VRAM Nvidia graphics card is sufficient for operation, and the text moderation performance is comparable to the GPT-OSS-Safeguard-20B, which has about 7 times the parameters.
