Key Takeaways:
Chinese AI models now command nearly two-thirds of global large language model invocations, a milestone that signals a structural shift in the competitive balance between US and Chinese AI ecosystems.
Xiaomi's MiMo-V2.5 ranked first globally on OpenRouter in July with 32.3 trillion monthly token calls, while weekly token volume exceeded 10 trillion, the platform's data show. The model, designed for autonomous agent scenarios with support for 1 million-token context windows and multimodal understanding of text, images, video and audio, has seen weekly token calls surge 616% from 1.46 trillion in May to 10.46 trillion.
"The Chinese models, I find them to be pretty close on code and research," said Curt Meinhold, founder of digital legacy platform LilyList and a Greensboro, North Carolina-based technology executive. "If I can pay a handful of cents per million output tokens versus $30 or $40 or $50, then it's good enough."
The top five models by OpenRouter invocation volume in July were all Chinese: Xiaomi's MiMo-V2.5, DeepSeek's V4 Flash, Tencent's Hy3 (free tier), MiniMax's M3 and Z.AI's GLM 5.2. As of July 26, Chinese models held 63.5% of global invocation share over the prior 28 days, according to OpenRouter, a platform that aggregates usage across hundreds of models. The data reflects a broader trend: US investment bank Goldman Sachs wrote in a July research report that Chinese AI models are reaching a "critical stage" for wide adoption, particularly as agentic AI usage — which requires models to autonomously conduct multistep tasks — drives demand for cost-effective alternatives.
Cost Advantage Drives Adoption
The pricing gap between Chinese and US frontier models is the primary catalyst. DeepSeek's V4, previewed in April, and Alibaba's Qwen3.8 Max, previewed in July, offer inference costs measured in cents per million output tokens, compared with $30 to $50 for Anthropic's Fable or Mythos models. Moonshot's Kimi K3, launched in July, generated more than 930,000 downloads in its first week, a 200% increase from the prior week, according to Sensor Tower. US downloads alone jumped 387% to about 86,000.
Raffi Krikorian, chief technology officer at Mozilla, said he switched to Kimi K3 for daily tasks within days of its launch, calling it "snappier" than Anthropic's Claude Fable. He had previously used Z.AI's GLM-5.2 for routine work including calendar management, documents and email.
Most Chinese AI models are open-source, while frontier models from Anthropic and OpenAI remain closed-source. "The open frontier is becoming increasingly Chinese-built," Krikorian said. A group of US technology firms including Microsoft, Meta and Nvidia signed an open letter in July backing open AI models, even as the Trump administration weighs restricting Chinese AI access in the US.
Limits and the Investment Picture
Despite the usage dominance, Chinese models still lag US leaders across full-range capabilities, said Anastasios Angelopoulos, co-founder and CEO of Arena, a platform for evaluating AI systems. US export controls on advanced AI chips remain a constraint, though they have also created openings: when the Trump administration restricted exports of Anthropic's Fable and Mythos models for more than two weeks in June, Z.AI released its GLM-5.2 model shortly after. "Restricting an American model can immediately create an opening for a Chinese competitor," Angelopoulos said.
The divergence between technological achievement and stock performance is stark. Xiaomi shares rose 1.32% to HKD29.06 on the day of the OpenRouter data release, while MiniMax-W tumbled 11.76% and Z.AI dropped 17.80%, suggesting investors are reassessing monetization paths despite strong adoption metrics. Z.AI reported revenue of 724 million yuan ($107 million) in 2025, up 132%, but its net loss widened to 4.7 billion yuan ($694 million), highlighting the sustainability challenge facing Chinese AI startups that are spending heavily on compute infrastructure.
For investors, the takeaway is that Chinese AI models have crossed a threshold from curiosity to mainstream adoption in cost-sensitive workloads. The question is whether the usage share can translate into revenue before the funding runs out.
This article is for informational purposes only and does not constitute investment advice.
