According to VisionOne Beating's monitoring, in the Big Model Invocation Leaderboard released by the model aggregation platform OpenRouter, Chinese big models have shown a historic "dominance" trend. Among them, the DeepSeek V4 Flash released on April 24 made its debut on the monthly leaderboard after only one month online with a direct invocation volume of 7.99T (trillion) tokens.
Based on OpenRouter's latest LLM invocation volume monthly leaderboard, the top ten rankings are as follows:
1. DeepSeek V4 Flash: 7.99T tokens
2. Hy3 preview (Tencent Hybird 3.0 Preview): 7T tokens
3. Claude Sonnet 4.6: 6.65T tokens
4. Claude Opus 4.7: 6.07T tokens
5. Hy3 preview (free): 5.83T tokens
6. Kimi K2.6: 5.45T tokens
7. Gemini 3 Flash Preview: 4.57T tokens
8. DeepSeek V3.2: 4.06T tokens
9. DeepSeek V4 Pro: 3.4T tokens
10. MiniMax M2.7: 2.95T tokens
In addition to DeepSeek holding three seats (V4 Flash, V3.2, V4 Pro), Tencent Hybird (Hy3 Preview and Free Version totaling 12.83T tokens), Moon's Dark Side (Kimi K2.6), and MiniMax (M2.7) also demonstrated strong performance. Among mainstream Western models, only Anthropic and Google held three seats, showing passivity in both total volume and momentum.
This dramatic change in the landscape is core to the dual breakthrough of Chinese big models in "ultimate cost-effectiveness" and "long-text processing." The lightweight MoE architecture represented by DeepSeek V4 Flash perfectly fits the rigid cost reduction requirements of self-programming intelligent agents such as Cursor and Claude Code in high-frequency and large-context reading, becoming the default choice for global developers in the intelligent agent scene.
