Alibaba says 3 billion Qwen downloads; Hugging Face counted 2.05 billion

Alibaba said this week that its Qwen models have passed 3 billion downloads, which it presented as evidence that Qwen is the most used open-weight family in the world ^[1]. The Hugging Face report published the day before, and cited in support of the claim, counts 2,045 million Qwen downloads this year, or 2,061 million if you include every repository ^[3].

The 3 billion number did not come from the report. According to The Next Web, it came from Alibaba's own emailed statement ^[2]. The gap is 955 million downloads ^[1]. That is 32 percent of Alibaba's figure, or 47 percent more than Hugging Face recorded, depending on which end you measure from ^[2][3]. TNW characterised it as roughly a third larger ^[4].

The derivatives claim is looser still. Alibaba says the Qwen ecosystem has produced more than 300,000 derivative models ^[4]; Hugging Face counts 151,448 Qwen-based derivatives on its Hub ^[5]. The lab's number is about double the platform's ^[5], a difference of roughly 148,000 models ^[6].

For a team choosing a base model, none of this changes the ranking. Hugging Face's 151,448 is still 2.6 times Meta's entire footprint on the Hub ^[6], which implies something on the order of 58,000 for Meta ^[7], and TNW notes the gap does not change who is first ^[7]. The report's own language is about workflow rather than volume: "Qwen has become part of the default workflow for developers deciding what models to fine-tune and deploy" ^[8].

What changes is where you get your inputs. Hugging Face is explicit that its download counts measure activity inside its own hub and do not capture API usage, private deployments or models distributed elsewhere, and Alibaba runs its own distribution platform in ModelScope ^[9]. That cuts both ways: the Hub number is a floor on real usage, not a ceiling, but it is the only one of these figures with a stated method. A vendor total that arrives by email has no method attached, and this is not the first time. TNW reported earlier that Alibaba called Qwen3.8 the world's second-best model without showing proof ^[10].

The underlying strategy is reach, and it is legible without the rounding. Alibaba has open-sourced more than 460 models under permissive licences and pushes them through its cloud into Southeast Asia and Africa ^[11], while also saying it wants to start charging the heaviest users of those open models ^[12]. Derivative counts are the more useful signal for procurement anyway: they tell you whether other teams have already done fine-tuning, quantisation and tooling work you can inherit. At 151,448 that answer is still yes.

Two things to watch. Meta and Nvidia have both shipped new open models in recent weeks, and Nvidia has been explicit that it is chasing China on open weights ^[13], so the derivative gap is the number to re-check rather than the download headline. The larger risk to Qwen's position is domestic: Beijing has been weighing curbs on overseas access to the country's best models ^[14], which would remove the distribution that produced these figures in the first place. If you are standardising on Qwen, plan an exit path now, and pull your adoption evidence from the Hub each quarter rather than from launch statements.