Kimi VL A3B Thinking is a compact visual-language model that applies extended reasoning to image understanding tasks. It uses a sparse mixture-of-experts architecture, activating only 3B parameters at inference time despite a larger total parameter count — keeping compute costs low while handling multi-step visual problems. It thinks before answering, which means responses take longer but tend to be more deliberate on complex diagrams, charts, or math-heavy images.