Mistral AI announced the release of Shieldstral, a 3‑billion‑parameter open‑weight multimodal safety classifier, under the Apache 2.0 license. The model runs on a single NVIDIA GPU with 16 GB of memory and provides policy‑adaptive moderation without requiring retraining. The release follows research that showed a smaller model can match or exceed the performance of larger guard models when trained on diverse, instruction‑based data. Traditional safety systems embed fixed categories and need retraining to adapt to new policy definitions, while Shieldstral accepts plain‑language policy questions at inference time, allowing a single checkpoint to evaluate text, images, or text‑image pairs. Training used a collection of publicly available and synthetic datasets that were converted to a uniform instruction‑query‑document format. To improve policy discrimination, the team generated contrastive pairs that differ by a single policy violation, enabling the model to learn the boundaries of each policy rather than memorize fixed labels. Image safety data were expanded with general image collections and a vision‑language reranker to reduce labeling noise. The final model combines three checkpoints — a base instruction model, a policy‑tuned checkpoint, and a public‑data calibrated checkpoint — through LoRA fine‑tuning and SLERP merging, preserving instruction following while adding moderation capability. Benchmarks indicate that Shieldstral matches or outperforms open‑source guard models up to seven times larger on text safety, refusal detection, and multimodal moderation tasks. The model produces a calibrated probability score from a single forward pass, allowing developers to set confidence thresholds or rank results. Mistral intends to extend multilingual support, improve handling of longer documents, and encourage community development of the model, which is now publicly available for download.