DeepSeek has just officially rolled out the latest iteration of its large-scale model, the V4.1 Flash-1017. At a pivotal moment in the evolution of China's artificial intelligence large models, this new version has taken the lead in implementing upgrades, boasting significantly enhanced performance. Prior to this, public data had already provided a glimpse into its technical prowess. This time around, DeepSeek has, for the first time, independently unveiled the Pareto frontier curve, offering a clear visual representation of the model's overall efficiency and cost-effectiveness. The V4.1 Flash-1017 shines on this curve, with its cost per unit task being a mere one-twentieth to one-hundredth of that of comparable models such as Fable and Astra, thus delivering a cost-performance advantage that is forty to fifty times greater.

Although the 1017 version retains the original API pricing structure, it has undergone profound optimizations in terms of Token usage efficiency. By employing task-aware compression, dynamic context pruning, and adaptive decoding strategies, it has substantially cut down the actual Token consumption for a wide range of request types. Real-world tests reveal that Token usage has plummeted by over 50% in most scenarios compared to the previous version, resulting in a notable decrease in the total cost of ownership (TCO).