Introducing Shieldstral.

Mistral AI unveiled Shieldstral, a 3-billion-parameter open-weights multimodal safety classifier designed to evaluate both text and images against user-defined policies at inference time. Unlike traditional guardrail models that require retraining for new contexts, Shieldstral accepts plain-language policy questions—such as whether content promotes violence or is suitable for minors—and returns calibrated safety scores without additional training. The model operates efficiently on a single 16GB NVIDIA GPU and is released under the Apache 2.0 license, making it accessible for deployment across diverse applications. Mistral AI positions Shieldstral as part of its contribution to the Open Secure AI Alliance, alongside NVIDIA and other organizations.

Shieldstral unifies multiple safety tasks—prompt classification, response moderation, refusal detection, and toxicity detection—into a single problem by framing content moderation as a policy-adaptive question-answering task. At inference, the model processes a policy query and content (text, images, or both) to output a continuous safety score derived from yes and no logits, eliminating the need for discrete labels. This approach allows policies to be dynamically adjusted at deployment without retraining, addressing the limitations of fixed taxonomies that often fail to align with specific product or audience requirements. The model’s performance matches or exceeds open guard models up to seven times its size across text safety, refusal detection, policy adaptability, and multimodal benchmarks.

The development of Shieldstral involved addressing challenges in data heterogeneity, policy discrimination, and image safety grounding. Public safety datasets often use inconsistent taxonomies and labels, so Mistral AI standardized datasets into a unified instruction–query–document format while calibrating strictness levels for different sources. To prevent memorization of predefined policies, the team trained the model on contrastive policy pairs, teaching it to distinguish between closely related policies rather than relying on fixed labels. For image safety, where data is scarce, the team supplemented moderation datasets with general-purpose image datasets, augmented queries, and used a vision–language reranker to filter mislabeled data and reduce hallucinations.

Shieldstral was built end-to-end on Mistral AI’s Forge platform, which handled distributed training, infrastructure, and evaluation metrics. The final model combines complementary checkpoints—one calibrated on public data, another fine-tuned for policy discrimination, and a base instruct model—via SLERP merging to balance calibration, adaptability, and instruction-following. Mistral AI emphasizes that Shieldstral represents a shift toward context-adaptive moderation, with ongoing work focused on expanding multilingual coverage, improving robustness for longer documents, and broadening multimodal safety capabilities. The company also invites community contributions and highlights open positions for those interested in advancing AI safety.