DeepSeek's New Model Reduces Token Consumption by Over 50%

Odaily Planet Daily reports that DeepSeek has officially launched the new version of its large model, V4.1 Flash-1017. At a time when domestic AI large models are generally in a critical phase of iteration, this version has taken the lead in completing its upgrade, achieving a significant leap in overall performance. Previously, publicly available data had already demonstrated its technical metrics. This time, DeepSeek has, for the first time, independently released the Pareto frontier curve, visually presenting the model's comprehensive performance in terms of efficiency and cost dimensions. The V4.1 Flash-1017 has reached an industry-leading level on this curve: the comprehensive cost per unit task is only one-twentieth to one-hundredth that of similar models like Fable and Astra, with a cost-performance advantage exceeding 40 to 50 times. Although the newly released 1017 version maintains the original API pricing structure, it has undergone a deep reconstruction in terms of token usage efficiency. Through task-aware compression, dynamic context trimming, and adaptive decoding strategies, it has significantly reduced the actual token consumption for various types of requests. Tests show that in most scenarios, token usage has decreased by more than 50% compared to the previous version, and the corresponding TCO (Total Cost of Ownership) level is expected to see a substantial new round of reductions. (Sina Finance)