Abstract:
According to the latest data from OpenRouter, the total number of global AI large model calls last week (August 31 to September 6) was 115 trillion Tokens, a month-on-month increase of 1.77%. Among the large AI models on the list, the weekly use volume of China's large AI models reached 56.72 trillion Tokens, a month-on-month increase of 2.83%; during the same period, the weekly use volume of large American AI models was 16.54 trillion Tokens, a month-on-month decrease of 3.1%. The number of weekly calls for large models in China has surpassed that of the United States for 19 consecutive weeks, ranking first in the world.
"Daily Economic News" noticed that last week, among the top five global calls, there were four large Chinese AI models. Among them, Tencent Hunyuan Hy4 preview ranked first, with weekly calls reaching 14.7 trillion Tokens, a month-on-month increase of 379%.
Hy4 preview was officially released and open sourced on August 28, with a total parameter of 770B, an activation parameter of 49B, and a context length exceeding 1M. It focuses on optimizing Agent, Coding and productivity scenarios to enhance the ability to complete complex tasks in real business scenarios.
On September 1, Tencent Hunyuan team announced the launch of the Hy4 preview lightweight version model, compressing the weight of Hy4 preview from 1.5TB to approximately 214GB, further lowering the local deployment threshold. According to reports, the long text understanding ability of the compressed model is almost the same as that of the original model, and the multiple rounds of long context retrieval are basically at the same level as the original version.
GPT-5.6 Luna ranked second, with weekly calls reaching 12.9 trillion Tokens, a month-on-month increase of 66%; GLM-5.3 Flash rose to third place, with a weekly call volume of 12.4 trillion Tokens, a month-on-month increase of 101%; DeepSeek-V4-Flash-0731 (the official version of DeepSeek-V4-Flash) ranked fourth, with a weekly call volume of 12.4 trillion Tokens; DeepSeek-V4-Flash-0423 (a preview version of DeepSeek-V4-Flash) ranked fifth, with a weekly call volume of 5.19 trillion Tokens.
After nearly a month, MiniMax M3 returned to the list, ranking sixth, with weekly calls reaching 5.02 trillion Tokens, a month-on-month increase of 95%. MiniMax previously announced that from August 24 to September 6, developers can use models such as MiniMax M3 for free through GMI Cloud and OpenRouter.
MiniMax M3 was released and open sourced in June this year. It is MiniMax’s flagship model for coding, agents and native multi-modal scenarios. The model conducts multi-modal mixed training of text, images and videos from the early stage of training, and supports millions of Token-level contexts.
In addition, just last week, HUMAIN, an artificial intelligence company under the Saudi Public Investment Fund (PIF), launched the first large Arabic model HUMAIN M3, which is based on MiniMax M3 for Arabic language and localization ability training. Unlike in the past, Chinese large models mainly entered overseas markets through APIs or application products. In this cooperation, MiniMax M3 directly became the basic model for the development of overseas local models.
It is worth noting that Xiaomi MiMo-V2.5, which ranked third in the previous week, and Gemini 3.7 Flash, which ranked ninth, fell off the list.
Comments