Classifies text, images, and video, extracts structured values, and locates objects in a single forward pass with calibrated probabilities. Built for low-latency decisions in workflows, not open-ended conversation.