Chinese Models Top Global AI Token Usage

🔑 Enhanced Key Takeaways

• •Chinese models captured 61% of total token volume on OpenRouter, with MiniMax M2.5 at 2.45 trillion tokens, Kimi K2.5 at 1.21 trillion, and Zhipu's GLM-5 at 780 billion.[2][3]
• •MiniMax M2.5 launched on February 13, 2026, as the world’s first production-grade model natively designed for agent scenarios, achieving 3.07 trillion tokens in its first seven days and a 197% week-over-week surge.[2]
• •Moonshot AI's Kimi K2.5 generated revenue in under 20 days post-launch exceeding its entire 2025 total, with overseas revenue surpassing domestic for the first time, fueled by global paid subscribers and API usage.[2]
• •Pricing advantage is key: MiniMax M2.5 and GLM-5 at $0.30 per million input tokens vs. Claude Opus 4.6 at $5.00, making them 16.7 times cheaper.[2]

🛠️ Technical Deep Dive

• •MiniMax M2.5: 230B parameters, 205K context window; excels in real-world software engineering with SWE-bench Verified 80.2 (highest), Multi-SWE-Bench 51.3, BrowseComp 76.3; trained for agent scenarios, office tools (Word, Excel, PowerPoint), context switching, and token-efficient planning.[1][4]
• •Kimi K2.5: 1T total parameters (32B active per token, MoE architecture), 262K context window; native multimodal with visual coding and self-directed agent swarm; standout benchmarks include HumanEval 99.0, MMLU 92.0, MMLU-Pro 87.1, LiveCodeBench 85.0, AIME 2025 96.1, GPQA Diamond 87.6, MATH-500 98.0, Chatbot Arena 1447, IFEval 94.0; continued pretraining on 15T mixed visual/text tokens.[1][4]

🔮 Future ImplicationsAI analysis grounded in cited sources

⏳ Timeline