Mistral AI has released Shieldstral, a 3-billion-parameter open-weights safety classifier the company says matches or outperforms guardrail models up to seven times its size, under an Apache 2.0 licence.
Unlike traditional guardrail models that rely on a fixed taxonomy of harm categories baked into their weights, Mistral said Shieldstral accepts plain-language safety policies at inference time, letting a single checkpoint adapt to new moderation contexts without retraining. The model evaluates text and images through a unified question-answering format, reading a policy, a yes/no query, and the content itself, then returns a calibrated safety score rather than a discrete label.
Mistral said it evaluated Shieldstral against guardrail models up to seven times its size across four benchmark categories, text safety, refusal detection, policy adaptability, and multimodal safety, using evaluation samples held out from training. The company said the model runs on a single 16GB NVIDIA GPU and was trained on a mix of real and synthetic data spanning multiple existing safety taxonomies, unified into one training format. Mistral said it used contrastive training pairs, generated by rewriting safe text to violate specific policies, to teach the model to distinguish between similar but distinct policy boundaries, rather than memorising fixed categories.
Shieldstral was built on Forge, Mistral's platform for training and evaluating custom models. Mistral said it is an inaugural member of the Open Secure AI Alliance, alongside NVIDIA and other organisations, and plans further work on multilingual coverage and broader multimodal safety.
