Shieldstral fits multimodal moderation in 16GB

Mistral released Shieldstral, a 3-billion-parameter multimodal safety classifier that fits in 16 GB of VRAM and scored 97.7% F1 on the VLGuard benchmark. The Apache 2.0 model card says operators can express a text-or-image moderation policy as a natural-language yes/no question at inference time, then receive a continuous score from one forward pass without retraining the checkpoint. For teams building the request guardrails that sit beside identity and tool policy, Shieldstral makes policy iteration cheap enough to test locally—but Mistral’s own benchmark should be the start of a product-specific eval, not the deployment verdict.