Select Language:
Chinese large language models continue to dominate global AI usage, maintaining their lead over U.S. competitors for the third consecutive week, according to rankings released by OpenRouter. Last week, five of the top nine most-used models worldwide were Chinese, the leaderboard showed. The top four included Xiaomi’s MiMo-V2-Pro, Shanghai Jieyue Xingchen Intelligent Technology’s Step 3.5 Flash, MiniMax’s M2.5, and DeepSeek’s V3.2. Additionally, Z.ai’s GLM5 Turbo secured the sixth position.
Overall usage reached 7.359 trillion tokens, marking a 57 percent increase from 4.69 trillion tokens the previous week. In contrast, U.S.-based models like Claude Sonnet 4.6 and Claude Opus 4.6 from Anthropic, along with Google’s Gemini 3 Flash Preview and Gemini 2.5 Flash, accounted for a total of 3.536 trillion tokens, reflecting a 7.3 percent rise from 3.294 trillion tokens the week before.
Last year, Chinese open-source models experienced the highest global download volumes. Their popularity is partly driven by being freely available, but also because they enable developers worldwide to customize and fine-tune these models. This approach has lowered barriers to AI adoption globally and contributed to the technology becoming a public resource, according to Joseph Tsai, chairman of a prominent Chinese tech conglomerate, speaking at the China Development Forum.
Further evidence of Chinese AI models’ growing international influence emerged last week when a U.S.-based AI coding company was found to have based its latest release on a Chinese startup’s model. The firm’s newest product, Cursor Composer 2, was quickly identified to be essentially Kimi 2.5, developed earlier this year by Moonshot AI from China. Users noted the similarities online, with one remarking that Composer 2 is just Kimi 2.5 with added reinforcement learning. Tesla CEO Elon Musk commented publicly, confirming the resemblance by stating, “Yep, it’s Kimi 2.5.”
The company behind Cursor promptly acknowledged the connection. Founder Adam Sanger posted on X, apologizing for not mentioning the Kimi model initially and promising to rectify this oversight in future versions.




