Qwen at 3 billion downloads: the open-weight default is now Chinese

Alibaba says its open-weight Qwen models have passed 3 billion global downloads in the past six months, overtaking Meta, Alphabet and its Chinese peers ^[1]. Hugging Face's state of open models report, published Aug. 14, put Google at 418 million downloads and Meta at 227 million in 2026 ^[3], which means the layer most teams fine-tune and deploy on is no longer American-made.

The gap is not marginal. On those figures Qwen's total is roughly 7.2 times Google's and 13.2 times Meta's ^[14][15], and about 4.7 times the two of them combined ^[16]. Alibaba also says it has open-sourced more than 460 models with an ecosystem of over 300,000 derivatives ^[2], which works out to roughly 650 derivatives per released checkpoint ^[17]. Hugging Face's own language is the more telling data point for anyone doing procurement: Qwen "has become part of the default workflow for developers deciding what models to fine-tune and deploy," and its download base makes it "one of the largest foundations of the open AI ecosystem" ^[7][6].

Two caveats on the arithmetic. The 460-model and 300,000-derivative counts come from Alibaba's own emailed statement, while the Google and Meta numbers are attributed to Hugging Face ^[2][3]. And the windows do not obviously match: Alibaba's 3 billion is described as covering six months, the rival figures as 2026 totals ^[1][3]. A like-for-like recount could compress the multiple. It would not reverse the ordering.

Why it stuck is more interesting than the number. Open models are building blocks that can be downloaded and customised, so adoption is a reasonable proxy for what developers are actually choosing ^[4]. Chinese labs including Alibaba have pushed models that are capable, relatively cheap and easy to adapt, and download and derivative counts have become one measure of standing in the US-China race ^[5]. Qwen, Moonshot AI and DeepSeek are all working to close the distance to closed US systems from OpenAI and Anthropic ^[9]. Alibaba is winning against those domestic rivals too, gaining on DeepSeek, Moonshot's Kimi and MiniMax as well as on US models ^[8].

Distribution is the part rivals cannot copy quickly. Alibaba pushes Qwen through its cloud platform to enterprise customers in Southeast Asia and Africa, reach that many competitors lack ^[11], and a broad family compounds: more adopters produce more derivatives, which draw more adopters ^[12]. That is a lock-in mechanic dressed as generosity. Once your tuning scripts, evals and serving stack assume a Qwen tokenizer and license, switching is a project, not a config change.

Export controls have not changed the trajectory. Restrictions on chips and AI systems, including the brief ban on overseas access to Anthropic's Fable 5 model this summer, do not appear to be slowing Chinese competitors ^[10]. The response so far has been product, not policy: Meta and Nvidia have both shipped new open models in recent weeks as the fight for developers intensifies ^[13].

Watch three things. Whether Hugging Face's next report normalises the time windows, which would tell you how much of the 7x is real and how much is calendar. Whether Meta and Nvidia's new releases move derivative counts rather than download counts, since derivatives are the stickier signal ^[13][2]. And whether Western enterprises that have quietly standardised on Qwen weights start being asked, by customers or regulators, to justify the dependency.