In the post-ChatGPT era, Chinese artificial intelligence was long presented as an open counterpoint, but in practice, Chinese models lagged behind in advanced development levels. However, a recent report from Hugging Face's Spring changes this picture. Open-weight models developed in China now account for 41 percent of total downloads on the world's largest model hub, surpassing American-origin models for the first time. The total number of downloads has exceeded 10 billion.
On the OpenRouter platform, four models with the largest token volumes over the past week are Chinese. Kimi K3, released by Moonshot AI on July 27th and boasting a total of 2.8 trillion parameters, is the largest open-weight model ever released. Three days later, Alibaba introduced Qwen3.8-Max, a base model with 2.4 trillion parameters that also has an open-weight plan.
The economic and political situation is straightforward: US export controls make major Nvidia chips unavailable to Chinese labs, forcing them to strive for algorithmic efficiency that would otherwise be merely desirable elsewhere. Investors support open distribution because open downloads represent the cheapest barrier against closed competitors. Andreessen Horowitz partner Martin Casado noted last year that about 80 percent of Silicon Valley startups seeking funding were building their products based on open Chinese models. Hugging Face CEO Clément Delangue asserts that China will likely catch up to leading American labs by the end of 2026.
The technical landscape is also transforming. Ion Stoica, a professor at the University of California, Berkeley, who co-authored Spark and Ray, told Chinese press that the gap between Chinese open models and global closed advanced models has narrowed from six to nine months to two to three months. Nvidia founder Jensen Huang publicly supported the open approach.
Venture capital is following this trend. DeepSeek completed its first external round worth over 50 billion yuan. Moonshot AI closed a Series F round, valuing it at $35 billion, and began a Series G round ahead of its IPO. Currently, hot money is looking not for whoever releases the biggest model fastest, but for the open model that will become the foundation for all others, much like Linux became the foundation of cloud technologies.
State industrial policy acts as a tailwind rather than the main driver. The fifteenth five-year plan prioritized the construction of open systems nationally. The national community of open AI now comprises over 11 million users and more than 70,000 models. The infrastructure exists because work was already underway; policy is only accelerating adoption, not invention itself.
The next stage is global expansion. Chinese open models have become the standard starting point for developers in Southeast Asia, Africa, and Latin America. While the United States still leads in raw advanced capabilities, the question for the next two years is whether the rest of the world will build its AI on Chinese open weights, just as it built cloud technologies on Linux.
Initially starting as fan art, a meme featuring DeepSeek founder Liang Wenfen in the guise of a 'cool guy' in a blue suit quickly gained significant resonance on Chinese social media. This meme served as a catalyst for the release of DeepSeek V4 Flash.
The DeepSeek V4 Flash model achieved a score of 7.1 trillion tokens in the weekly global OpenRouter ranking and instantly became the most used model globally. As a result, DeepSeek secured two spots in the world's top five, with four positions held by Chinese models, while GPT-5.6 Luna dropped to sixth place.
Economic indicators show a significant advantage. One day before the release of V4 Flash, OpenAI reduced the price of GPT-5.6 Luna by 80 percent, setting the cost at $0.2 per million input tokens and $1.2 per million output tokens. DeepSeek V4 Flash was released at a price equivalent to approximately $0.14 for input and $0.28 for output, while offering a 98 percent caching discount compared to the industry standard of 90 percent. The independent evaluator Artificial Analysis determined the cost of V4 Flash for a typical agent task to be about 60 percent lower than that of Luna. OpenCode recorded a caching rate of 96 percent, which lowered the average session cost to $0.06.
The performance gap is no longer substantial. DeepSeek's internal tests show that V4 Flash surpasses V4 Pro Preview and GLM 5.2, only trailing Opus-4.8. Furthermore, Artificial Analysis assigned it an intelligence index score of 50, whereas Luna received 51. On the VulcanBench benchmark, Bold Metrics CTO Morgan Linton characterized V4 Flash as 'exceptionally stable under moderate load,' noting its superiority over Claude Fable and GPT-4.5. Approximately half of the traffic directed to DeepSeek on OpenRouter now comes from developers in the US and Europe, indicating that this phenomenon extends beyond the domestic market.
The focus on agent development is deliberate. V4 Flash utilizes the V4 Flash Preview architecture and focuses on agent capabilities after training. Team member Cui Tianyi publicly hired engineers for a new group, Harness—a system that wraps the model with context, tools, task state, and feedback, using open GitHub projects as a selection filter. In May, the DeepSeek-TUI agent, developed by independent American developer Hunter Boone, gathered over 11 thousand stars on GitHub in just a few days. Boone's open letter, titled 'whale bros, I'm the American who builds DeepSeek-TUI,' became a cultural event that swept across the Pacific. Since international developers support these models through their choices, China's open LLMs are no longer just a curiosity; they have become the baseline price-to-performance ratio that all others must now surpass.
Chinese artificial intelligence models accounted for 63.5% of global queries to large language models (LLMs), surpassing the US, which holds 35.5% of the market as of July 2026. Xiaomi MiMo-V2.5 became the leader in global calls on the OpenRouter platform, processing 10.5 trillion tokens per week and 31.2 trillion per month.
According to data, for the period ending July 26, Chinese models demonstrated a significant increase in their market share compared to early 2025, when American models dominated global AI workloads. This shift is attributed both to the rapid improvement in the quality of Chinese models and aggressive pricing strategies that make them more accessible to developers worldwide.
The rise in popularity of Xiaomi MiMo-V2.5 was particularly noticeable. Since the public beta launch on April 23, 2026, the weekly token volume increased from 1.46 trillion in May to 10.46 trillion in July, representing a 616% increase over two months. This model is positioned as a universal foundation model designed for the agent era, supporting million-level context windows and full understanding of all modalities: text, images, video, audio, and speech.
MiMo-V2.5 has quickly gained recognition among developers due to competitive pricing and reliable performance in agent-oriented tasks. The data has several important implications for the global AI industry. Firstly, inference volume has become a more indicative measure of real AI adoption than traditional quality benchmarks. Secondly, the scale of inference indicates that the battle for market share in AI has shifted from training capabilities to inference economics—the ability to serve billions of tokens at a sustainable unit cost.
Intra-China competition fosters price reduction and quality improvement, creating a positive cycle. Chinese APIs have become the standard for cost-sensitive tasks. Although Xiaomi entered the market later than companies originally focused on AI, it managed to scale faster thanks to its hardware and software integration and its developer ecosystem.
