Shieldstral
Shieldstral is a 3B open-weights multimodal safety classifier that takes plain-language policies at inference time, with no retraining.
Frames moderation as a policy-conditioned yes/no question over text, images or both, covering prompts, responses and refusal detection. Mistral says it matches or beats guardrail models up to 7x larger and runs on a single 16GB GPU. Apache 2.0.
- Date
- Tuesday, 4 August 2026
- Lab
- Mistral AI
- Kind
- open-weights
- Access
- open weights
Figures
| Measure | Value | Measured by |
|---|---|---|
| Parameters | 3B runs on one 16GB GPU | company |
Sources
This record was checked against its sources on 6 October 2026. How we check