Mistral releases Shieldstral: 3B multimodal safety classifier runs on a single GPU

MistralAI · x · 2026-08-05

Mistral AI has introduced Shieldstral, a multimodal safety classifier. Despite having only 3 billion parameters, the model matches or outperforms models nearly 7x its size on text safety benchmarks and sets a new state-of-the-art in multimodal safety classification.

Shieldstral formulates content moderation as a binary question-answering task, unifying diverse moderation tasks into a single yes/no problem to consolidate heterogeneous safety datasets under one framework. Furthermore, it boasts high efficiency, running on a single 16GB NVIDIA GPU, and allows enterprises to customize safety policies.

Related event: Mistral Releases Shieldstral: A 3B Open-Source Content Safety Model(3 posts)→

Original post →

More from Models

Models channel →