Mistral has released Shieldstral, a 3-billion-parameter AI safety model that runs on a single 16GB GPU. The company says it matches or outperforms open guardrail models up to seven times larger and sets a new state of the art in multimodal moderation.
The open-weight model can assess text, images, and combined text-and-image content. Mistral released it under the Apache 2.0 license.
Safety Policies Without Retraining
Most guardrail models rely on fixed harm categories and require retraining when developers change their safety policies.
Shieldstral instead lets developers describe a policy as a plain-language question during inference. The model then returns a calibrated safety score without retraining.
Each request includes an instruction describing the context, a yes-or-no safety question, and the content being assessed. The model uses the “yes” and “no” probabilities to generate a continuous score in a single forward pass.
This approach allows one model to handle prompt classification, response moderation, refusal detection, toxicity detection, and image safety.
How Mistral Trained It
Mistral trained Shieldstral on real and synthetic datasets with different labels and safety categories. It converted them into one format and created contrastive examples to help the model distinguish between similar policies.
The team also added general image datasets to improve visual moderation and filtered image-question pairs to reduce incorrect labels. It combined several LoRA-trained checkpoints using SLERP and built the model through its Forge training platform.
Mistral plans to improve Shieldstral’s multilingual support, long-document performance, and broader multimodal safety capabilities.
The post Mistral’s New AI Runs on One 16GB GPU, Beats Models 7x Bigger appeared first on ProPakistani.