“KIMI K3 failed the "Lava Lamp benchmark," struggling with shape, motion, and visual quality, while GPT-5.6 produced a cleaner result.”