Mistral Launches Shieldstral, a 3B Open-Weights Safety Classifier for Text and Images
Mistral AI has released Shieldstral, a 3-billion-parameter open-weights safety classifier that moderates both text and images. Rather than scoring content against a fixed set of harm categories, Shieldstral treats moderation as a question-answering task: developers write their policy as a plain-language question at inference time, and the model returns a calibrated safety score in a single forward pass.
Details
- Question-answering framework: Shieldstral takes three inputs — an instruction with evaluation context, a yes/no policy question, and the content to judge (a prompt, a response, or an image) — and returns a probability by reading the logits for “yes” and “no”
- No retraining needed: Because policies are expressed as natural-language questions rather than baked-in categories, teams can adapt moderation rules on the fly without fine-tuning a new model
- Unified text and image support: A single model interface handles text-only, image-only, and combined text-image moderation
- Performance: Mistral says Shieldstral matches or exceeds guard models up to 7x its size across text safety, refusal detection, policy adaptability, and multimodal safety benchmarks, all evaluated on held-out data
- Efficient footprint: Runs on a single 16GB NVIDIA GPU
- Built on Forge: Trained using contrastive policy pairs and LoRA checkpoints merged via SLERP, built entirely on Mistral’s Forge training platform
- Open Secure AI Alliance: Shieldstral ships as a founding release under the newly announced Open Secure AI Alliance, alongside NVIDIA and other partners
How to try it
- Shieldstral-1.0-3B is released under the Apache 2.0 license and available for download on Hugging Face at
mistralai/Shieldstral-1.0-3B - Because it runs on a single 16GB GPU, it can be self-hosted for on-premises or air-gapped content moderation pipelines