“In a one-million-token context setting, DeepSeek-V4-Pro requires 27% of the single-token inference FLOPs compared to DeepSeek-V3.2.”