OpenRouter data for the week of September 14–20 shows global large-model calls reached 129 trillion tokens, with Chinese models at 67.46 trillion tokens—4.7x the United States and a 21st consecutive week at number one worldwide. DeepSeek's V4.1-Flash, released September 10, topped the chart at 15.8 trillion tokens (up 219% week on week), followed by Zhipu GLM-5.3-Flash and Tencent Hunyuan Hy4 preview.
In parallel, testimony to the US Congress noted that Chinese developers have accumulated 3.2 billion downloads of open-weight AI models on Hugging Face—double the US total—capturing 80% of OpenRouter traffic as of September 2026. Chinese open models now trail American closed frontier labs by only two to five months, while American open releases lag six to nine months behind.
The trend underscores that open-weight models have become the operating foundation for global downstream enterprise and developer applications, with price-performance "high value for money" routing global inference traffic toward Chinese models at a quantifiable pace.
Source: OpenRouter / AI Breaking Wire / Artificial Analysis
