Chinese Models Lead Weekly Token Volume for 20th Straight Week as DeepSeek V4.1 Flash Hits No. 6

A calculation by the Chinese business daily National Business Daily, based on the latest OpenRouter data, puts global large-model token consumption at 127 trillion tokens in the week of Sept. 7 to Sept. 13, up 10.43 percent week over week. Chinese models accounted for 61.17 trillion tokens, up 7.85 percent, while U.S. models accounted for 21.76 trillion, up 31.56 percent.

Four of the top five models by weekly token volume were Chinese. Tencent's Hunyuan Hy4 preview ranked second with 16.8 trillion tokens, up 15 percent. The report notes that Hy4 preview was officially released and open-sourced on Aug. 28 as a model with 770 billion total parameters and 49 billion active parameters, a context window beyond 1 million tokens, and tuning for agent, coding and productivity workloads.

Zhipu's GLM 5.3 Flash placed third with 11.9 trillion tokens, down 4 percent. DeepSeek V4 Flash 0731, the official release version of V4 Flash, was fourth with 11.6 trillion, down 6 percent, and Xiaomi's MiMo-V2.5 climbed to fifth with 7.77 trillion, up 230 percent.

The week's biggest newcomer was DeepSeek V4.1 Flash. Released on Sept. 10, it reached sixth place within three days of launch with 4.94 trillion tokens. According to the report, V4.1 Flash uses a new Causal-Encoder-Decoder architecture and is a mixture-of-experts model with 552 billion total parameters.

DeepSeek also cut API pricing for V4.1 Flash. Per million tokens, cached input costs 0.02 yuan, uncached input 1 yuan and output 4 yuan during off-peak hours, while peak-hour prices are 0.04 yuan, 2 yuan and 8 yuan. The new prices took effect at 12:00 on Sept. 10, 2026. Compared with V4 Flash at the same hours, cached input is down 60 percent, uncached input about 33.3 percent and output about 11.1 percent.

DeepSeek says V4.1 Flash now surpasses V4 Pro across performance, cost, speed and total task time. Meanwhile MiniMax M3, sixth the previous week, and Zhipu GLM 5.3, ninth, dropped off the list, a sign that turnover inside the ranking has accelerated.

The growth rates tell a subtler story. China's base is far larger but grew 7.85 percent, while U.S. models, at roughly a third of the volume, grew 31.56 percent, suggesting U.S. demand is accelerating again. Twenty weeks at the top does not mean the race has stopped.

Three things to watch: whether V4.1 Flash keeps climbing after its price cut, whether GLM 5.3 Flash and Tencent's Hunyuan Hy4 preview hold their positions, and whether a pricing war spreads to more vendors as agent workloads push token consumption higher.