# Mistral AI Open-Sources Multimodal Content Moderation Model Shieldstral

- Category: Flash (快讯)
- Published: 2026-08-05T11:52:00.000Z
- Source date: 2026-08-05
- HTML: https://www.88z.ai/en/news/flash-20260805-c8fb20ea91b24f19
- Markdown: https://www.88z.ai/en/news/flash-20260805-c8fb20ea91b24f19.md
- Other language: https://www.88z.ai/news/flash-20260805-c8fb20ea91b24f19.md

Mistral AI has released Shieldstral, an open-source multimodal content moderation model supporting 12 languages and runnable on a single 16GB GPU. The model achieves state-of-the-art performance in multimodal moderation, rivaling competitors seven times its size. It innovatively inputs moderation policies as natural language questions, outputting calibrated safety scores via a yes/no Q&A mechanism, suitable for prompt classification, response moderation, and toxicity detection.

## At a glance

- Shieldstral supports 12 languages, runs on a single 16GB GPU, and matches the performance of competitors seven times its size.
- It uses natural language question inputs and a yes/no Q&A mechanism to output calibrated safety scores for various moderation tasks.

## The update

Mistral AI has officially open-sourced Shieldstral, a multimodal content moderation model. Supporting 12 languages, including Chinese, the model is hardware-efficient and can be deployed on a single GPU with 16GB of VRAM. In terms of performance, Shieldstral achieves state-of-the-art results in multimodal moderation, matching the capabilities of competitors that are seven times larger in scale. Architecturally, the model innovatively converts moderation policies into natural language questions for input. It then outputs calibrated safety scores through a yes/no question-and-answer mechanism. This design allows it to flexibly handle various security tasks, such as prompt classification, generated content moderation, and toxicity detection, providing developers with an efficient and precise content safety solution.

## Sources

- [IT之家](https://www.ithome.com/0/985/886.htm) (2026-08-05)

This article is an AI-assisted synthesis by 88Z AI based on the public sources below. Product claims and availability are attributed to their publishers. It does not represent independent testing or first-hand reporting.