How Chinese AI Companies Captured the Global Open-Weight Market

Here is how that advantage shows up in the data.

Download dominance: Qwen overtakes Llama

The first clear signal came from Hugging Face, the largest open-source AI model platform. In September 2025, Alibaba's Qwen model family surpassed Meta's Llama to become the most downloaded LLM family on the platform .

Over the broader period from August 2024 to August 2025, Chinese open-weight models captured 17% of all global downloads on Hugging Face, surpassing the US (15.8%) for the first time, according to a joint study by MIT and Hugging Face . Among large language models specifically, Chinese-developed models accounted for 41% of all LLM downloads on the platform during that period .

By mid-2026, cumulative worldwide downloads of Chinese models had reached over 100 million, according to Hugging Face data cited by China Daily .

Token volume on OpenRouter: from <2% to outright majority

While downloads measure developer interest, token consumption measures actual usage. On OpenRouter — the largest neutral router of LLM traffic, processing north of 20 trillion tokens per week by early 2026 — the data shows a complete reversal .

Chinese-origin models went from less than 2% of weekly token volume to the majority share in about 18 months . By May 2026, Chinese open-weight models accounted for roughly 61% of all tokens consumed on the platform . In July 2026, Chinese models held all five top spots on OpenRouter's global usage rankings for the first time, with Xiaomi's MiMo-V2.5 ranking first by token volume, followed by models from DeepSeek, MiniMax, Alibaba's Qwen family, and Moonshot's Kimi .

US enterprise adoption surged

Perhaps the most striking figure for US observers: by mid-2026, Chinese AI models captured a weekly peak of 46% of US enterprise API token volume on OpenRouter, compared with 35.7% for US-origin models . That was up from under 2% a year earlier. On every week since February 8, 2026, Chinese-origin models have accounted for at least 30% of enterprise token volume on the platform .

This shift was driven by a combination of competitive performance and drastically lower cost. For example, Moonshot's Kimi K2.5 came close to Anthropic's Claude Opus on some early benchmarks but cost roughly one-seventh the price . DeepSeek's V3.1 charges 2 yuan ($0.28) per million input tokens and 8 yuan ($1.10) per million output tokens .

Domestic Chinese consumption exploded

The data also reflects massive growth inside China. China's National Data Administration reported that daily domestic token consumption surpassed 140 trillion in March 2026, a more than 1,000-fold increase from 100 billion in early 2024 . ByteDance's AI assistant Doubao alone went from consuming about 100 billion tokens per day in May 2024 to over 120 trillion per day by March 2026 — a roughly 1,000-fold surge in two years .

The engine: a rapid release cadence at permissive licenses

What enabled all of this? The data points to three structural factors:

1. Pace of releases. Between late April and mid-June 2026, four frontier-competitive open-weight models shipped within weeks of each other from Chinese labs: DeepSeek, Alibaba's Qwen, Moonshot's Kimi, and Zhipu's GLM .
2. Permissive licensing. Models are released under MIT, Apache 2.0, and modified MIT licenses, making them freely downloadable, modifiable, and deployable at scale .
3. Cost advantage. Chinese labs charge far less than US competitors for API access, a gap that has persisted even as performance has closed .

The competitive moat, analysts conclude, is less about a single breakthrough model and more about this self-reinforcing ecosystem — a loop that, as of mid-2026, has given Chinese companies the lead in the open-weight AI layer.