Alibaba Qwen: 3 billion downloads and a shift in the balance of power in open AI
Alibaba's Qwen open-model ecosystem has reached an impressive milestone: downloads have exceeded 3 billion over the past six months. The Chinese corporation has publicly released more than 460 neural networks, based on which developers worldwide have created about 300,000 derivative solutions. This is not just statistics, but a marker of a fundamental shift in the industry.
Hugging Face Data: Dominance Without Qualification
Analytics from the Hugging Face platform over the first seven months of this year paint an even more telling picture. In this ecosystem alone, Qwen models were downloaded 2.05 billion times, and the number of forks and derivative repositories reached 151,448. For comparison, Google's figure stands at a modest 82,506. It is important to understand that Hugging Face statistics only account for activity within its own platform, excluding API calls and private corporate deployments.
Google's total downloads for 2026 are estimated at 418 million, and Meta's at 227 million. These figures, however, should not be perceived as a direct indicator of market share or commercial success—the report's authors themselves caution against such simplified interpretations. Nevertheless, Qwen's growth rate speaks volumes: the number of derivative repositories increases by 180–210 daily. The community's contribution is also telling: out of 28,531 GGUF conversions of these models, only 54 were created by Alibaba itself.
"Qwen has become a standard element of the workflow for developers choosing which model to fine-tune and deploy," note Hugging Face analysts.
The Secret to Success: A Scaling Strategy
Qwen's superiority is explained by three key factors: regular updates to the model lineup, coverage of all scales—from sub-billion versions to the giant Qwen3.8-Max with 2.4 trillion parameters—and the open Apache 2.0 license, which removes restrictions on commercial use. The decisive factor turned out to be precisely the wide range: models with fewer than 1 billion parameters account for 83% of all historical platform downloads, while neural networks over 100 billion account for only 1%. Competitors focused on giant LLMs are clearly losing: Moonshot AI, which releases almost no models smaller than 70 billion, gathered only 37 million downloads over the year—roughly 55 times fewer than Qwen. In the local deployment segment, the gap is also enormous: Qwen GGUF builds are downloaded 39.6 million times per month, compared to 20.8 million for Gemma and 7.5 million for Llama.
A New Balance of Power in Open AI
The report records a tectonic shift: in almost every month, the largest open model from China surpassed all American releases in size. The parameter ceiling for Chinese models ranged from 754 billion to 2.78 trillion, while for US competitors it did not exceed 130 billion in five out of seven months. Licensing policies also differ dramatically: 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. In the US, only 29% fall under open licenses, 41% under proprietary terms, and 30% have no license specified at all.
Notably, in the US, the drivers of open models are not AI labs but chip manufacturers: AMD and Nvidia have each released more than 200 repositories, noticeably outpacing Google and Meta.
My comment: This data is vivid confirmation that the "openness as a weapon" strategy works. Alibaba is not just catching up with Western competitors but is reshaping the market by betting on mass adoption and a developer ecosystem. Western labs should think twice: control over infrastructure matters, but without open models that developers choose, leadership in AI could turn out to be a Pyrrhic victory.