Alibaba Says Qwen Passed 3 Billion Downloads — Ahead of Meta and Google, by Alibaba’s Own Math
- 3B+ Hugging Face downloads Alibaba's Qwen model family has logged, per a Bloomberg report citing a Hugging Face study relayed via an Alibaba company statement — Bloomberg, August 15, 2026
- 1.2B Meta's own reported all-time cumulative Llama downloads — nearly six times the 227 million figure Bloomberg's comparison credits to Meta — Meta, LlamaCon, April 2025
- 900M Google's own reported all-time cumulative Gemma downloads — more than double the 418 million figure in Bloomberg's comparison — Google Blog, July 2026
- 153.6M Downloads Qwen logged in February 2026 alone — more than its next eight competitors' monthly totals combined — The ATOM Project
Alibaba’s Qwen family of open-weight AI models has passed 3 billion downloads on Hugging Face, the machine-learning hub where developers pull the underlying files that power chatbots, coding assistants, and enterprise AI tools. Bloomberg’s Saritha Rai reported the figure this week, citing a Hugging Face study relayed through an Alibaba company statement, and framed it as Qwen outpacing the comparable figures for Meta’s Llama (227 million) and Google’s Gemma (418 million). Alibaba also says it has released more than 460 Qwen model variants, spawning over 300,000 community-built derivatives.
That comparison needs a caveat before anything else: Meta itself reported 1.2 billion all-time cumulative Llama downloads back in April 2025, and Google has separately reported roughly 900 million all-time cumulative Gemma downloads as of July 2026 — both larger, on a different basis, than the 227 million and 418 million figures Bloomberg’s comparison uses. The likeliest explanation is that Bloomberg’s numbers cover calendar-year 2026 downloads only, not lifetime totals, but neither Bloomberg’s report nor the outlets reprinting it say so. Readers should not conclude from this comparison that Qwen has more total downloads, ever, than Llama or Gemma. It may not.
The more defensible story is the one underneath the download count: how fast Chinese open-weight models, Qwen chief among them, have become infrastructure that American cloud providers and enterprises now build on by default — arriving the same month a White House AI adviser warned Congress is fumbling the response, and the same window in which a federal standards agency flagged security risks in Chinese open-weight models generally.
Bloomberg’s comparison (Aug. 15, 2026): Qwen 3 billion+, Meta 227 million, Google 418 million — sourced to a Hugging Face report cited via an Alibaba company statement.
Each company’s own all-time total: Meta 1.2 billion Llama downloads (April 2025); Google roughly 900 million Gemma downloads (July 2026).
Bloomberg’s own figures for Meta and Google are far smaller than each company’s self-reported lifetime total — a strong sign the comparison is scoped to 2026 downloads only, not all-time downloads. Neither Bloomberg’s piece nor its secondary pickups say so.
None of this makes the 3-billion figure false — Hugging Face download counts are real and trackable, and Qwen’s growth is not in dispute, as the next section shows. What’s misleading is the framing: a calendar-year snapshot presented as proof Qwen has “passed” Meta and Google, when both companies’ own published totals suggest the opposite may be true on a lifetime basis.
It’s also worth naming who did the framing. The comparison didn’t originate with Hugging Face publishing a headline finding; it came from Alibaba handing a data point to a reporter as a company statement. That doesn’t make the download count fabricated — but it makes the “beats Meta and Google” framing Alibaba’s marketing conclusion, not Hugging Face’s.
Set the scope-mismatch problem aside, and Qwen’s growth curve on Hugging Face is still striking on its own terms. The ATOM Project, an independent open-weight-model tracking initiative founded by AI2 researcher Nathan Lambert, found Qwen overtook Llama as the single most-downloaded open-weight model family on Hugging Face around September 2025. Qwen crossed 1 billion cumulative downloads on January 21, 2026, and in February 2026 alone it logged 153.6 million downloads — more than the combined monthly totals of its next eight competitors, including Meta, DeepSeek, OpenAI, Mistral, Nvidia, Zhipu, Moonshot, and MiniMax.
Alibaba marked the 1-billion milestone publicly in January:
1 Billion Downloads: A New Milestone for Qwen — Hugging Face data shows Qwen became the first open-source LLM to surpass 200K derivative models — has now exceeded 1B total downloads, averaging 1.1M/day, surpassing Meta's Llama as the world's No. 1 open-source large model.
Seven months later, Qwen’s downloads had tripled to more than 3 billion under Bloomberg’s count — growth that would be remarkable under any accounting basis, calendar-year or lifetime.
There’s a second problem with the 3-billion figure beyond the scope mismatch: this reporting could not independently locate or read the specific Hugging Face report Bloomberg cites. The only Hugging Face report publicly accessible at the time of writing was an earlier “State of Open Source AI” edition from the spring of 2026 — and that report does not contain a download-comparison table matching the numbers in Bloomberg’s piece. Hugging Face itself has not published a headline finding declaring Qwen ahead of Llama and Gemma on downloads; the comparison traces back to Alibaba’s own statement citing the data.
Nearly every outlet covering the story is reprinting Bloomberg’s wire text rather than independently verifying the numbers — standard for a same-day wire story, but worth flagging. The South China Morning Post, reporting separately, put Qwen’s share of global open-source downloads at “over 50%” — a figure broadly consistent with the trajectory the ATOM Project has independently tracked since 2025, even if the exact comparison basis differs from Bloomberg’s.
The download count, however imprecisely framed, points at a real and under-covered story: American cloud infrastructure now distributes Qwen as a first-class option. Amazon Web Services, Microsoft, Databricks, and Google’s own Vertex AI platform all make Qwen models directly available to their enterprise customers — the same hyperscalers that also sell access to Llama and Gemma now routinely offer a Chinese-developed alternative alongside them.
Hugging Face CEO Clem Delangue told CNBC on August 3 that China is “winning the AI race” on open models specifically, predicting the country could reach the technological frontier by the end of 2026 or in 2027. Delangue runs the platform where the download counts at the center of this story are generated, giving him a direct vantage point on adoption trends that neither Alibaba nor its American competitors control.
“China is winning the AI race — at least when it comes to open models.”
Clem Delangue · CEO, Hugging Face · CNBC interview, August 3, 2026
The same dynamic prompted White House AI adviser David Sacks to warn, in a July 17 post about a different Chinese model — Moonshot AI’s Kimi K3, which had just taken the top spot on the Frontend Code Arena benchmark — that U.S. policymakers risk losing the broader race over open-weight AI.
This is concerning. For the first time, a Chinese model Kimi K3 has taken #1 on the Frontend Code Arena and is scoring at or near the frontier on other benchmarks. Meanwhile America is tying itself in knots: politicians and bureaucrats are banning new data centers, piling on [state regulations]... This is how you lose the AI race.
Sacks’ post concerned Kimi K3, not Qwen — but the warning applies to the broader pattern Qwen’s download numbers illustrate. That pattern carries a documented downside, too: one analysis tied to NIST’s Center for AI Standards and Innovation examined DeepSeek’s open-weight models specifically — not Qwen — and found AI agents built on them roughly 12 times more likely to follow malicious hijack instructions than comparable American models. No equivalent finding about Qwen has been published, but the study underscores that widespread enterprise adoption of any Chinese open-weight model family is arriving faster than independent security review of it.
Qwen’s download numbers sit inside a larger financial picture at Alibaba that is considerably messier than a milestone headline suggests. Alibaba Cloud Intelligence Group, the unit that develops Qwen, reported revenue of RMB 41.6 billion for the quarter ended March 31, 2026 — its most recent reported quarter as of publication — up 38% year over year, with AI-related products posting triple-digit growth for an eleventh consecutive quarter. Alibaba Group’s total revenue for the same quarter was RMB 243.4 billion, up about 3% year over year — but non-GAAP diluted earnings per share fell 95% year over year, as heavy capital spending on AI infrastructure compressed margins even while cloud and AI revenue climbed.
Qwen’s technical leadership has also been in flux. Junyang Lin, the researcher widely credited as Qwen’s technical lead, departed Alibaba in March 2026 amid a reported internal dispute over compute allocation and restructuring with Alibaba Group CEO Eddie Wu. Alibaba Cloud CTO Zhou Jingren was named to continue leading Tongyi Lab, the research group that builds Qwen, in Lin’s place.
None of that undercuts the download growth. It does mean the “beats Meta and Google” headline arrives from a company burning margin to fund the infrastructure behind that growth, while absorbing a leadership departure on the team that built the product — context a one-line comparison leaves out.
Qwen’s download growth is real and independently corroborated — but the specific claim that it has “passed” Meta and Google traces to a single comparison Alibaba handed to one reporter, measured against figures far smaller than what Meta and Google themselves report as their all-time totals. The number worth watching isn’t 3 billion. It’s that AWS, Microsoft, Databricks, and Google Vertex AI now all distribute a Chinese-developed model by default, that Hugging Face’s own CEO says China is winning the open-model race, and that the security review of these models is visibly lagging their adoption.
