Behind Deepseek's 85% speed: big model rolls, costs War

In June 2026, Leung Wenfeng published a DSpark paper, which resulted in an 85 per cent increase in the speed at which services are generated under real traffic on the DeepSeek-V4 line. This is not simply an upgrade of hardware, but rather a waste of computing power resulting from the elimination of invalid verifications through confidence-based scheduling. Large model competition is shifting from parameter scale to system-level engineering games that reason efficiency and cost。