DeepSeek V4 Flash đạt 72,2 tỷ token trong sử dụng hàng tuần, dẫn đầu bảng xếp hạng mô hình AI toàn cầu

The Huoxing Finance report, based on data from the multi-model aggregation platform OpenRouter, shows that from July 27 to August 2, DeepSeek V4 Flash ranked first on the global weekly leaderboard for AI large model API calls, with 7.22 trillion tokens processed. According to the open-source platform OpenCode, this model alone processed 8 trillion tokens on August 1. Four of the top five models are Chinese: Xiaomi MiMo-V2.5 (5.1 trillion), Tencent Hunyuan Hy3 (5.01 trillion), and another DeepSeek V4 Flash (3.45 trillion) ranked second to fourth, respectively, while OpenAI GPT-5.6 Luna (2.99 trillion, up 738% week-over-week) ranked fifth. Chinese large models collectively processed 28.13 trillion tokens last week, marking the 14th consecutive week they have surpassed the U.S. and maintained the top position globally. Previously highly watched Kimi K3 dropped out of the top ten this week, placing 12th. Notably, free models are capturing an increasing share of traffic, with some models recording week-over-week growth exceeding 50%. Analysts note that Chinese models have shifted from price-based competition to relying on inference performance and stable service capabilities; however, OpenAI, Anthropic, and others continue rapid iteration, accelerating the pace of internal competition, where product iteration speed will determine user retention.