# Shieldstral 1.0

A compact multimodal moderation model for prompt moderation, response moderation, prompt-response pair classification, refusal detection, and safety filtering across text and image inputs. It uses natural-language policy questions and returns a yes or no classification.

- Status: Public Preview
- Type: Labs, Specialist
- Release date: 2026-08-04
- Version: 1.0
- Context length: 32k
- Parameters: 3.8B total and active
- Input: text, image
- Output: text
- Licenses: Apache 2.0
- Open weights: [Weights](https://huggingface.co/mistralai/Shieldstral-1.0), 3.8B parameters, Apache 2.0
