Mistral AI has released 'Shieldstral,' a small moderation AI capable of analyzing text and images using natural language rules.

On August 4, 2026, French AI company Mistral AI released 'Shieldstral 1.0 3B,' an AI model that determines whether text and images meet specified safety standards. It boasts 3 billion parameters and is characterized by its ability to specify service-specific safety standards using natural language.

Introducing Shieldstral. | Mistral AI

🛡Introducing Shieldstral, Mistral's 3B open-weights model for content safety that can be deployed on-device 🧵 https://t.co/SXBTWIJO9p pic.twitter.com/Hhjw8o0gft

— Mistral AI (@MistralAI) August 4, 2026

In services that use AI-generated content, 'content moderation' is performed to determine whether it is appropriate to display or provide text entered by users or responses generated by the AI. For example, it detects content that promotes violence, discriminatory language, or text related to self-harm, and stops the response as needed.

On the other hand, the scope of what should be considered dangerous varies depending on the service; for example, content that is deemed safe in cybersecurity research tools may be considered harmful in mental health services. Traditional safety assessment models often have predetermined classifications of dangerous content, and changing the criteria sometimes required additional model training.

Shieldstral allows users to input criteria for what they want to assess, such as 'Does this promote violence?' or 'Is this image safe for minors to see?'. The model reads the text or image in question and quantifies its safety based on the probability of 'yes' or 'no.' The service provider can then hide the content or forward it for human review based on the numerical value.

In an evaluation of Mistral AI that examined whether it could adapt to new safety standards without additional training, Shieldstral performed comparably to larger models and outperformed several of its comparison targets.

It supports not only text but also images, and can determine whether a standalone image or an image with text is safe using the same mechanism. In an evaluation of image moderation by Mistral AI, Shieldstral recorded the highest overall score among the models compared.

Shieldstral can be used not only to inspect user input and AI responses, but also to determine whether the AI has appropriately rejected dangerous requests. It supports 12 languages, including English, Japanese, Chinese, and French, and runs on a single GPU with 16GB of VRAM.

The model weights are released under the Apache License 2.0 and can be downloaded from Hugging Face. Mistral AI states that it will continue to improve its multilingual support, enhance the stability of its long sentence recognition, and expand the scope of its moderation to include images.

mistralai/Shieldstral-1.0-3B · Hugging Face
https://huggingface.co/mistralai/Shieldstral-1.0-3B