Shieldstral
2026-08-04 · Publisher : Mistral AI · Domain : Artificial Intelligence
Mistral AI launches Shieldstral 1.0, a multimodal safety classifier with 3 billion parameters under the Apache 2.0 license, which applies policies written in natural language without retraining.
What changes
- Works as a binary question-answering system over policies provided at inference time.
- Returns calibrated safety scores for text and images.
- Matches or exceeds open guard models up to 7 times larger (text, refusals, policy adaptability, multimodal).
- Built with Mistral's Forge platform.
### Specifications - Parameters: 3B. Modalities: text and image. License: Apache 2.0 (open weights). - Runs on a single 16 GB NVIDIA GPU. - Availability: Hugging Face (mistralai/Shieldstral-1.0-3B); also listed in the documentation catalog. - API identifier, context and cutoff: not disclosed. ### Pricing Pricing not disclosed at launch (downloadable open weights). ### Benchmarks The post claims performance comparable or superior to models 7 times larger, with no figures included here. Scores published by the vendor, not detailed in the pages consulted. ### Key takeaways An open, lightweight guardrail, adaptable simply by writing a policy, suited to companies that need to self-host their safety filters.
Links
Tags : LLM, Mistral, Shieldstral, Sécurité, Open Source