ISLAMABAD – French artificial intelligence company Mistral has released Shieldstral, a 3-billion-parameter AI safety model designed to moderate text, images, and combined multimodal content while operating on a single 16GB graphics processing unit (GPU).
According to the company, Shieldstral delivers performance comparable to or better than open-source guardrail models that are up to seven times larger. Mistral also claims the model establishes a new benchmark for multimodal AI safety.
Unlike conventional guardrail models that depend on predefined harm categories and require retraining whenever safety policies change, Shieldstral allows developers to create customised safety rules during inference.
Read more: How to apply for Australia’s visa as a skilled worker?
Developers only need to provide a plain-language instruction describing the context alongside a simple yes-or-no safety question. The model then evaluates the submitted content and returns a calibrated safety score without requiring additional training.
What’s the scroing process
The scoring process relies on the probabilities assigned to “yes” and “no” responses, enabling Shieldstral to generate continuous safety scores in a single forward pass. This approach allows one model to perform several moderation tasks, including prompt classification, response moderation, refusal detection, toxicity detection, and image safety analysis.
Mistral said Shieldstral was trained using a combination of real-world and synthetic datasets covering multiple safety categories. The company standardised these datasets into a unified format before generating contrastive examples to improve the model’s ability to distinguish between closely related safety policies.
Read more: Is Jean Grey going to be the next MJ?
To strengthen multimodal moderation, Mistral incorporated general image datasets and filtered image-question pairs to minimise incorrect labels. The company also combined multiple LoRA-trained checkpoints using the SLERP merging technique and trained the final model through its Forge training platform.
Shieldstral has been released under the Apache 2.0 licence, allowing developers and organisations to use, modify, and distribute the model freely.
Read more: Active Matrix Cube sparks Gen Z trend focused on individuality
Key points:
- Mistral has introduced Shieldstral, a 3-billion-parameter AI safety model.
- The model runs on a single 16GB GPU despite outperforming some much larger guardrail models.
- Shieldstral supports moderation of text, images, and combined text-image content.
- Developers can define custom safety policies without retraining the model.
- The model is available under the Apache 2.0 open-source licence.
- Mistral plans to improve multilingual capabilities and long-document performance in future updates.
