Qwen Hits 3 Billion Downloads: Alibaba's Open Models Now Outpace Meta and Google Combined

On August 15, 2026, Bloomberg reported a milestone that would have sounded absurd two years ago: Alibaba’s Qwen family of open-weight AI models has accumulated more than 3 billion downloads in the past six months, making it the most-downloaded open model family in the world — and putting it far ahead of Meta and Google, the two companies that essentially invented modern open model publishing.

The numbers come from Alibaba, cross-referenced against Hugging Face’s “State of Open Models: Summer 2026 Observations” report published August 14. The contrast is stark. In 2026 to date, Google’s open models have drawn roughly 418 million downloads and Meta’s about 227 million — combined, barely a fifth of Qwen’s trailing-six-month figure. Qwen has open-sourced more than 460 models, and its ecosystem has spawned over 300,000 derivative models built by third parties.

For a sense of scale: the Hugging Face Hub itself hosts about 2.96 million public model repositories. One Chinese model family now anchors roughly one in twenty of them.

The Hugging Face report behind the headline

The download milestone is the headline, but the deeper story is in the biannual report from Hugging Face’s policy team (Adina Yakefu, Apolinário, and Irene Solaiman), which worked through data from January to August 2026. Several of its findings explain how Qwen got here — and why the gap is likely structural, not cyclical.

China owns the frontier end of open weights. In almost every month of 2026, the largest open model from a Chinese lab was bigger than anything released by an American lab of its own. China’s monthly ceiling ran between 754B and 2.78 trillion parameters; America’s stayed under 130B in five of seven months, with the only exceptions being NVIDIA’s Nemotron 3 Ultra (561B) and Thinking Machines Lab’s Inkling (952B). Moonshot, MiniMax, Xiaomi, and Z.ai publish almost nothing below 70B. Tencent and Alibaba’s Qwen cover the whole range from under 1B upward — and that full-spectrum strategy is precisely what correlates with adoption.

Qwen has become the community’s base model. Qwen-based models now account for 151,448 derivatives on the Hub — 2.6× Meta’s total footprint and 4.7× the Llama repositories specifically. Google follows with 82,506 derivatives. New Qwen derivatives are appearing at roughly 180–210 repositories per day. Notably, this was built by the community, not by Alibaba: of the 28,531 GGUF conversions of Qwen models, Alibaba itself published only 54.

Permissive licensing is the moat. Of 178 Chinese releases above 20B parameters this year, 59% carry Apache 2.0 and 22% MIT — and exactly none carry a non-commercial restriction. DeepSeek and Z.ai ship 700B-to-1.65T-parameter models under plain MIT. On the American side of the same size band, only 29% is Apache or MIT, 41% sits under custom terms, and 30% declares nothing at all. Chinese labs license their largest models more permissively than American labs license their smallest.

Attention ≠ adoption. The HF team’s sharpest methodological point: of the top 25 repositories by 2026 downloads and the top 25 by likes, exactly one appears in both. Not one model published in 2026 reaches the download top 25; thirteen of the twenty-five date from 2022. Downloads measure what production pipelines depend on; likes measure what the field is excited about. Qwen’s win is a win in the former category — the boring, sticky, infrastructure kind.

Scale in context: Moonshot vs Qwen strategies

The report quantifies the strategic divergence inside China itself. Moonshot’s frontier-only portfolio (Kimi models, nearly all above 70B) recorded 37M downloads over the period. Qwen’s full-spectrum strategy reached 2,045M across repositories with declared parameter counts (2,061M including all repositories) — about 55× more. The conclusion is unambiguous: coverage across sizes, not headline benchmark wins, is what converts into ecosystem capture.

This mirrors what happened in the token market. OpenRouter data shows Chinese models (DeepSeek, Moonshot, Z.ai’s GLM) now leading token volumes among developers, and this week the Financial Times reported that leading US model prices have fallen nearly 25% in a month — OpenAI cut GPT-5.6 Luna by 80% to $0.20/$1.20 per million tokens, and Anthropic priced Opus 5 at roughly half of Fable 5’s cost — partly in response to Chinese price pressure, with DoorDash, Siemens, and Airbnb all reportedly trialing Chinese models.

Why this matters

Three implications stand out.

First, the “who controls open AI” question has a new answer. Open model publishing was defined by Meta’s Llama and Google’s Gemma. In 2026, the two most prolific new open-model publishers in the US are AMD and NVIDIA — hardware vendors using models to sell chips — while Meta has moved toward closed flagships. The HF report’s phrasing is blunt: open source has moved from model labs to hardware and infrastructure companies in the US, while Chinese labs own the model layer.

Second, derivatives are the real adoption metric. Raw downloads can be inflated by CI pipelines and automated pulls. But 151,448 derivative repositories represent actual engineering investment by third parties — fine-tunes, quantizations, and rebuilds that lock developers into a model family’s architecture, tokenizer, and tooling. That is the kind of adoption that compounds.

Third, monetization is still an open question. As the HF report notes, whatever these releases are for, it is not license revenue — the return has to come from API and cloud business, hardware positioning, or ecosystem leverage. Alibaba’s bet is that being the default base layer for the world’s fine-tuners is worth more than any licensing income, the same playbook that made Linux the substrate of the cloud era.

Caveats worth noting

Download counts are an imperfect proxy. They don’t measure production usage, and the HF report itself warns that likes and downloads measure different things. Google and Meta also generate hundreds of millions of downloads annually through smaller and embedding models not counted in headline comparisons, and US participation in open-source AI is still growing in absolute terms. Meta’s pullback is a strategic choice, not a collapse.

But direction matters more than precision here. Six months, 3 billion pulls, 460+ models, 300,000+ derivatives, and a licensing posture more permissive than anything coming out of America — the open-weights center of gravity is no longer moving toward China. It has arrived.