China's Free AI Models Far More Advanced, American Products Left Behind
The open-source artificial intelligence (AI) model from China is becoming increasingly popular among global developers. Alibaba’s Qwen has become the most downloaded open-source AI model in the world, surpassing models from Meta and Google.
Citing China Daily, Qwen’s cumulative global downloads have surpassed 3 billion. The figure shows the increasingly widespread use of China’s open-source AI by developers in various countries.
A report on the development of open-source models published by the AI platform Hugging Face noted that Qwen models recorded around 2.045 billion downloads on the Hugging Face Hub throughout 2026. That figure is far above AI models from the United States, such as Google, which recorded around 418 million downloads, and Meta, around 227 million.
However, that figure only covers downloads via Hugging Face. The data does not yet include other platforms such as ModelScope in China. Therefore, Qwen’s global download count of more than 3 billion covers various platforms.
This development marks a shift in the global open-source AI competition. Chinese AI models are no longer competing only on performance and cost, but are also beginning to build developer ecosystems on a large scale.
Alibaba itself has released more than 460 Qwen models as open source. These models have in turn spawned more than 300,000 derivative models developed by the community.
On Hugging Face alone, developers have created 151,448 derivative models from Qwen. This number is 2.6 times more than derivative models from Meta and 4.7 times the number of repositories related to the Llama family.
Meanwhile, Google’s models have produced around 82,506 derivative models. The rapid adoption of Qwen shows the increasingly tight competition to determine which AI model will become the foundation for various applications.
Open-source models allow developers to perform fine-tuning for specific industry and language needs, carry out compression so they can run on personal computers or other devices, and integrate them into new applications. The impact is not only felt by the companies that create the models.
This difference means that open-source models often serve as the foundation for thousands of derivative products, even though the original developers do not operate those applications directly.
The Hugging Face report also shows a widening gap between Chinese and American AI labs in terms of the scale and licensing of open-source models.
The largest open-source models released by several leading Chinese AI labs in 2026 generally have a larger number of parameters compared to models released by American AI companies or labs.
According to the report, the largest parameter count of models from Chinese labs each month ranges from 754 billion to 2.78 trillion parameters. Meanwhile, the largest open-source models from the United States remain below 130 billion parameters in most months.
Even so, the number of parameters does not directly determine a model’s intelligence or efficiency. A mixture-of-experts (MoE) architecture, for example, can have hundreds of billions to trillions of parameters in total, but only activates a portion of the parameters for each request.
In addition to model size, Chinese developers also tend to use more permissive licences. These licences allow companies and developers to modify and commercialise the available AI models.
Of the 178 Chinese models with more than 20 billion parameters tracked in the report, 59% use the Apache 2.0 licence and 22% use the MIT licence.
The report notes that no model in that group imposes restrictions on use solely for non-commercial purposes.