Alibaba's open-source Qwen models have surpassed the 3 billion download mark: a phenomenon that is reshaping the AI market
Over the past six months, the ecosystem of open AI models Qwen from the Chinese corporation Alibaba has reached a colossal milestone — over 3 billion downloads. This is not just a number, but a marker of a tectonic shift in the industry: China has stopped catching up and has begun setting the pace.
The corporation has released more than 460 neural networks into the public domain, on the basis of which third-party developers have created about 300,000 derivative solutions. The scale is impressive, but even more interesting is the qualitative dynamics that I have been tracking in recent industry reports.
Numbers that speak louder than words
The independent platform Hugging Face, a leading hub for machine learning models, has published a semi-annual report on the state of open AI. According to this data, in just seven months of the current year, Qwen models were downloaded 2.05 billion times, and the number of their derivative repositories exceeded 151,000. For comparison: Google's figure is only 82,506. Over the entire year 2026, Google model downloads, according to the platform's estimates, reached 418 million, while Meta's were a modest 227 million.
It is important to understand that Hugging Face statistics only account for activity within its own ecosystem, excluding API requests and private corporate deployments. This means that the real scale of Qwen's distribution could be significantly higher. The report's authors rightly caution against interpreting these numbers as a direct indicator of market share, but the trend cannot be ignored.
Anatomy of success: why Qwen outpaces the giants
The analysis reveals three key factors behind Qwen's dominance. First, it is a regular update cycle that maintains community interest. Second, unprecedented coverage of scales — from compact versions with sub-billion parameter counts to the flagship Qwen3.8-Max with 2.4 trillion parameters. Third, and perhaps most importantly, is the Apache 2.0 license, which removes all restrictions on modification and commercial use.
The breadth of the model lineup proved to be a decisive factor. According to Hugging Face data, versions with fewer than 1 billion parameters account for 83% of all historical platform downloads, while giant neural networks with over 100 billion parameters occupy only 1%. This explains why laboratories focused exclusively on monstrous LLMs are falling behind. For example, Moonshot AI, which almost never releases models smaller than 70 billion, gathered only 37 million downloads over the year — roughly 55 times fewer than Qwen.
Balance of power and a new equilibrium in open AI
The report demonstrates a fundamental shift in the balance. In almost every month of the current year, the largest open model from China surpassed all American releases in size. The ceiling for Chinese models ranged from 754 billion to 2.78 trillion parameters, while for US competitors it did not exceed 130 billion in five out of seven months.
The licensing aspect is especially telling. In China, 59% of models with more than 20 billion parameters are released under Apache 2.0, 22% under MIT, and none have restrictions on commercial use. For American developers, the picture is diametrically opposite: 29% under open licenses, 41% under proprietary terms, and 30% without a specified license. Notably, in the US, the leaders in the number of new open models were not AI laboratories but chip manufacturers — AMD and Nvidia, each of which released more than 200 repositories.
My expert assessment: We are witnessing not just competition between models, but a clash of two philosophies. The Chinese strategy of "openness as a weapon" combined with aggressive scaling of small models creates a snowball effect that the US cannot yet stop. Control over infrastructure is a powerful trump card, but if the developer community ultimately pivots to open Chinese ecosystems, American technological leadership could be called into question already in the medium term.