Chinese open-source models account for 41% of downloads on the Hugging Face platform, surpassing the U.S. to become the global leader. Models from Tencent, Xiaomi, DeepSeek, and others occupy six of the top spots on OpenRouter’s most popular models. UBS reports that approximately 60% of enterprises have restricted AI spending by using “model routing” strategies to delegate simple tasks to low-cost Chinese open-source models, reserving high-end proprietary models for complex reasoning. Giants like Microsoft and AWS have already integrated Chinese open-source models, a trend reshaping the value distribution across the AI industry chain.Author and source: Caijing.com
On another front of the AI arms race, Chinese open-source models are quietly reshaping the competitive landscape through download statistics.
According to Hugging Face’s Spring 2026 report, Chinese open-source models account for 41% of the platform’s monthly and total downloads, surpassing U.S. models. Recently, this figure was again highlighted by Hugging Face CEO Clem Delangue, sparking widespread discussion on social media.
Hugging Face is currently the world’s largest open-source AI model hosting platform, hosting nearly 3 million public models and 1 million public datasets, with a new repository created every seven seconds on average. Half of the Fortune 500 companies have deployed their own private or open-source models on this platform.
Additionally, the top six most popular models on OpenRouter all come from Chinese institutions such as Tencent, Xiaomi, DeepSeek, MiniMax, and Zhipu.
Hugging Face CEO Clem Delangue (right)
Beyond the top six download rankings on OpenRouter, there has also been a significant shift in usage patterns.
Among the top six most popular models on the AI model invocation platform OpenRouter, all are from Chinese institutions—Tencent, Xiaomi, DeepSeek, MiniMax, and Zhipu are all included. Anthropic’s Claude Opus 4.7 ranks seventh.
Zhipu recently released the open-source model GLM-5.2, designed for agentic coding tasks, and it competes at the same level as Anthropic's latest model in relevant benchmark tests.
Infrastructure data further confirms this trend. According to Vercel platform data, in June 2026, open-source models handled nearly one-third of all AI requests on the platform, taking on a significant share of infrastructure-intensive workloads, while proprietary models gradually retreated to higher-cost, premium tiers.
Corporate shift: The core logic driving this transition from "renting" to "owning" is cost and control.
Delangue said: "If you're an AI or tech company, you wouldn't want to outsource your core capabilities to a black-box API you can't control." He noted that companies are reevaluating the value of owning their own AI after seeing the bills for scaling proprietary frontier models.
Microsoft CEO Satya Nadella also issued a similar warning against reliance on a single vendor, stating: "If learning flows in only one direction, economic value concentrates in the hands of those who own the learning infrastructure, rather than those who create knowledge." He advocates for decentralizing learning infrastructure to every enterprise.
The AI research report released on June 23, 2026, by Karl Keirstead’s team at UBS Securities found that approximately 60% of companies have restricted AI spending in some way, primarily by implementing guardrails on token usage.
Model routing becomes mainstream: As high-end models shift from default to luxury, companies are systematically reshaping the distribution of benefits across the AI industry chain.
According to UBS report, "model routing" has become a core technical strategy for token optimization—assigning different tasks to different models, where only complex reasoning, critical code, and long-context analysis invoke the most expensive models, while simple tasks are directed to lower-cost or even open-source Chinese models.
Price differences are the direct driver. For example, Anthropic’s Haiku 4.5 has an output price of $5 per million tokens, Opus 4.5-4.8 is $25, and the most advanced model reaches $50—resulting in a tenfold price difference between the entry-level and top-tier models.
Under this context, Chinese open-source models have entered the radar of corporate procurement. According to a case described in the UBS report, a major global bank has locally deployed Alibaba’s Qwen to balance the cost of using high-end models like Claude. The model catalog on AWS Bedrock now includes MiniMax, Kimi, Qwen, DeepSeek, and GLM; Microsoft Azure AI Foundry also offers access to DeepSeek.
Can open source determine the direction of the AI race? Delangue’s assessment is more direct. According to PodcastAlpha, Delangue believes open source is an accelerator for AI leadership: "If China continues to lead in open source, it could surpass the entire AI field within one to two years."
Dario Amodei, CEO of Anthropic, holds a different view. He believes there are risks in releasing open-source models with increasingly powerful capabilities, as they become difficult to control once released.
Delangue countered: "Locking it behind the doors of a few players doesn't make it secure." He added that transparency enables defenders to "patch cybersecurity risks already exposed in open-source models."