Mistral AI has released Shieldstral, a 3-billion parameter open-weights multimodal safety classifier designed for content moderation. Released under the Apache 2.0 license, the model aims to address limitations in traditional guardrail models, which often rely on fixed taxonomies of harm categories that require retraining to adapt to new contexts.
Shieldstral treats content moderation as a binary question-answering task. During inference, the model receives an evaluation context, a yes/no question, and the content to be judged (which can include text, images, or prompt-response pairs). By providing plain-language policies at the time of inference, the model can adapt to novel deployment contexts without the need for retraining. The model returns a continuous, calibrated safety score based on the probability of a ‘yes’ or ‘no’ response, allowing users to threshold or rank by confidence.
A key feature of the model is its efficiency and performance scale. According to Mistral, the 3B model matches or outperforms open guard models up to seven times its size across benchmarks for text safety, refusal detection, policy adaptability, and multimodal moderation. Due to its 3B parameter size, the model is capable of running on a single 16GB NVIDIA GPU.
To develop the model, Mistral addressed challenges in data heterogeneity by converting various datasets into a unified instruction–query–document format. The training process also involved using contrastive pairs—where text is engineered to violate one policy but not a similar one—to teach the model to distinguish specific policy boundaries rather than simply memorizing labels. For visual safety, which is limited by the scarcity of unsafe image data, Mistral supplemented moderation datasets with general-purpose image datasets and used a vision-language reranker to filter data.
Shieldstral’s release follows Mistral’s involvement as an inaugural member of the Open Secure AI Alliance alongside NVIDIA and other organizations.
Key facts
- Shieldstral is a 3B parameter open-weights multimodal safety classifier released under Apache 2.0.
- The model uses a question-answering task to enable policy adaptation at inference time without retraining.
- It can match or outperform models up to 7x its size in text safety and multimodal benchmarks.
- The model is efficient enough to run on a single 16GB NVIDIA GPU.
- It provides a calibrated safety score rather than a discrete label.