Alibaba's Qwen models just passed 3 billion downloads, more than Meta and Google combined

[Brad79]

Bloomberg is reporting that Alibaba's open weight Qwen models have racked up more than 3 billion global downloads over the past six months, which puts the company ahead of both Meta and Google combined to become the most downloaded AI model family in the world. The numbers come from a Hugging Face state of open models report published August 14, and they show Google at 418 million downloads and Meta at 227 million over the same period, which is not even close

Qwen has open sourced more than 460 individual models at this point, and developers have built over 300000 derivative versions on top of that base, which is the kind of ecosystem effect that tends to snowball once it gets going. Alibaba has also been pushing Qwen through its cloud platform into enterprise customers across Southeast Asia and Africa, giving it a distribution reach that a lot of its rivals simply do not have in those markets yet

Hugging Face called Qwen one of the largest foundations of the open AI ecosystem, and said it has basically become part of the default workflow for developers deciding what to fine tune and deploy. That is a pretty strong statement coming from the platform that hosts basically everyone's models, so it is not just Alibaba tooting its own horn here

Worth remembering that download counts measure adoption and reach rather than raw model quality, so this is not the same claim as Qwen being the smartest model out there. Still, US labs are clearly paying attention, since Meta and Nvidia have both pushed out new open model releases in just the past few weeks as the competition for developer mindshare heats up

Feels like a pretty significant marker in the open weight race even with that caveat. Developer adoption at this scale tends to compound over time rather than fade

Alibaba Ai Models Hit 3 Billion Downloads Passing Meta Googlewww.bloomberg.com

[Fraction]

3 billion versus Google and Meta combined not even hitting a billion is a genuinely lopsided number. Wonder how much of that gap comes down to pure pricing versus actual technical merit though, since Qwen has been notably cheap to run compared to a lot of Western alternatives

[CyberWarden49]

Download counts always feel like a slightly misleading metric to lead with.

A huge chunk of those downloads are probably duplicate pulls from CI pipelines and automated testing rather than real distinct developers picking Qwen over something else

[Joanne]

The enterprise distribution angle through Alibaba Cloud into Southeast Asia and Africa is the part that actually matters long term. Raw download numbers come and go, but if you get baked into someone's production infrastructure in an underserved market, that is real lasting leverage

[Cheeky Blake]

300000 derivative models built on top of Qwen is the number that actually impressed me here.

That is not the raw download count, that is genuine downstream build activity

[BillyBob]

Nvidia and Meta both dropping new open models within weeks of this report coming out does not feel like a coincidence. That reads like a direct competitive response even if nobody says so publicly

[Northern Ridge]

Genuinely curious whether this changes how Western enterprises think about their own AI vendor selection. A lot of companies have been quietly reluctant to build on Chinese models for supply chain and data governance reasons. I doubt a download milestone alone shifts that calculus much

[Hyperdrive71]

The framing of eclipsing Meta and Google combined is doing a lot of work in that headline.

Still impressive on its own terms, but worth remembering those two companies also gate plenty of their strongest models behind closed APIs rather than releasing weights at all, so the comparison is not perfectly apples to apples

[CollapseState]

Fair point in the article that downloads measure reach and not quality. I would still bet a good chunk of the developer base migrating this heavily is doing real comparative testing before switching. Not just following hype for its own sake

[DigitalNomad]

Is there any breakdown of where these downloads are actually coming from geographically. Or is it just a global aggregate number.

Would love some regional detail either way

[Dom9]

460 open sourced models under one family umbrella is a lot of surface area to maintain. Has to be a significant engineering and documentation effort behind the scenes to keep all of those usable and current