The most-watched public scoreboard of developer AI demand has flipped. Chinese-origin models — from Xiaomi, Alibaba, DeepSeek, MiniMax and Moonshot — now carry more than 45% of all traffic on OpenRouter, the model-routing marketplace, up from under 2% a year ago. Four of the platform's top five models by token volume are Chinese; OpenAI's share sits near 7.5%.
The leader
At the top is Xiaomi's MiMo-V2-Pro: a mixture-of-experts model with over 1 trillion total parameters (42 billion active), a 1-million-token context window, and heavy optimization for agentic work. It processes more than 4 trillion tokens a week on OpenRouter — roughly 21% of the platform and more than double Claude Sonnet 4.6 — at $1 per million input tokens and $3 output, a fraction of Western frontier pricing.
The 'quiet ambush'
MiMo-V2-Pro arrived in stealth: an anonymous model called Hunter Alpha appeared on OpenRouter on March 11 and climbed the leaderboard so fast the community assumed it was an unreleased DeepSeek V4. When Xiaomi revealed it on March 18, its stock jumped 5.8%. "I call this a quiet ambush — not because we planned it, but because the shift from Chat to Agent paradigm happened so fast," said Luo Fuli, Xiaomi's MiMo lead and a former DeepSeek researcher.
Price, not benchmarks
Western models still edge the benchmarks — MiMo-V2-Pro scores 78% on SWE-bench Verified against 80.8% for Claude Opus 4.6, and ranks eighth on Artificial Analysis's intelligence index — but the price gap is doing the work. GLM-5.2 runs $1.40/$4.40, and DeepSeek V4-Pro charges $0.44/$0.87. For high-volume agentic workloads where tokens are burned by the billion, 'good enough at a fifth of the price' is winning.
The caveats
OpenRouter measures developer and hobbyist routing, not enterprise procurement, where data-jurisdiction and compliance concerns still steer buyers Western. Reviewers also flag political-topic restrictions in Chinese models and inconsistent structured tool-calling. And MiMo-V2-Pro itself is closed-weight — a reminder that 'Chinese' and 'open' are overlapping but not identical categories, even as Qwen alone has spawned more than 100,000 derivative models on Hugging Face.
