According to OpenRouter data, global AI large model usage totaled 115 trillion tokens last week, up 1.77% week-over-week. China’s weekly large model usage reached 56.72 trillion tokens, up 2.83% week-over-week, marking the 19th consecutive week it has surpassed the U.S. and maintaining its position as the world’s largest market. Tencent HunYuan Hy4 preview led with 14.7 trillion tokens in weekly usage, a 379% week-over-week increase. Four of the top five models are from China. MiniMax M3 re-entered the rankings and was adopted by Saudi AI company HUMAIN as the foundational model for its Arabic language large model development, representing a significant breakthrough in the international export of China’s large model technology.Author and source: AIBase
According to the latest data from OpenRouter, the total global usage of AI large models last week (August 31 to September 6) reached 115 trillion tokens, representing a 1.77% increase week-over-week. Among these, Chinese AI large models accounted for 56.72 trillion tokens in weekly usage, up 2.83% week-over-week, while U.S. AI large models recorded 16.54 trillion tokens in weekly usage, a 3.1% decline over the same period. China’s weekly AI model usage has now exceeded that of the U.S. for nineteen consecutive weeks, maintaining its position as the global leader.
Four of the top five are from China, with Hunyuan Hy4 preview ranking first.
Four of the top five most-used AI large models globally last week were from China. Tencent’s HunYuan Hy4 preview ranked first, with a weekly usage of 14.7 trillion tokens, a 379% increase week-over-week. The model was officially released and open-sourced on August 28, featuring a total of 770 billion parameters and 49 billion activated parameters, with a context length exceeding 1 million tokens, optimized specifically for Agent, Coding, and productivity use cases. On September 1, the Tencent HunYuan team launched a lightweight version of Hy4 preview, reducing the model size from 1.5 TB to approximately 214 GB, further lowering the barrier to local deployment, while maintaining nearly equivalent long-text understanding capability compared to the original model.
Following closely are: GPT-5.6 Luna in second place with 12.9 trillion weekly API calls, a 66% increase week-over-week; Zhipu GLM-5.3 Flash rising to third place with 12.4 trillion weekly API calls, a 101% week-over-week increase; DeepSeek-V4-Flash-0731 (official version) in fourth place with 12.4 trillion weekly API calls; and DeepSeek-V4-Flash-0423 (preview version) in fifth place with 5.19 trillion weekly API calls.
MiniMax M3 re-enters the rankings, becoming a foundational model for overseas local model development.
After nearly a month, MiniMax M3 has reappeared on the leaderboard, ranking sixth with a weekly usage of 5.02 trillion tokens, a 95% increase week-over-week. MiniMax previously announced that developers could use MiniMax M3 and other models for free via GMI Cloud and OpenRouter from August 24 to September 6. The model, released and open-sourced in June this year, is designed for coding, agent, and native multimodal scenarios, undergoing multimodal training with text, images, and video from the earliest stages of training, and supporting context lengths of up to millions of tokens.
Last week, HUMAIN, an AI company under Saudi Arabia’s Public Investment Fund (PIF), launched its first Arabic large language model, HUMAIN M3, trained on MiniMax M3 to enhance Arabic language and localization capabilities. Unlike previous cases where Chinese large models primarily expanded overseas via APIs or application products, in this collaboration, MiniMax M3 directly serves as the foundational model for local AI model development abroad.
Notably, Xiaomi MiMo-V2.5, which ranked third the previous week, and Gemini 3.7 Flash, which ranked ninth, have dropped off the list.