Alibaba's Qwen models pass 3 billion downloads

Alibaba said its open-sourced Qwen model family — more than 460 released models and over 300,000 derivative models built by other developers — had been downloaded more than 3 billion times globally over the preceding six months, a figure Fortune and other outlets reported as putting it well ahead of Meta’s 227 million and Google’s 418 million open-model downloads for the year, according to data traced to a Hugging Face report, “State of Open Models: Summer 2026.”

That report itself gave a lower figure than Alibaba’s headline number. Hugging Face’s own tally, counting downloads recorded on its Hub, put Qwen’s 2026 total at roughly 2.05 billion — closer to 2.06 billion including every repository — and counted 151,448 derivative models rather than Alibaba’s claimed 300,000-plus. The Next Web noted the gap directly, describing Alibaba’s 3 billion figure as “about a third larger than the platform being cited actually recorded.” Hugging Face’s report also cautioned that Hub downloads capture only activity within its own ecosystem, not API usage, private deployments, or distribution through other channels — including Alibaba’s own ModelScope platform, which is not counted at all and would account for at least part of the discrepancy.

Even on Hugging Face’s more conservative count, Qwen’s download volume was reported as far ahead of any US open-weight family, continuing a pattern through 2026 in which Chinese labs — Qwen chief among them, alongside Moonshot and DeepSeek — published open weights for frontier-scale models more readily than US counterparts, whose flagship releases had increasingly stayed closed. Commentators framed the download counts, contested or not, as evidence that Qwen had become a default base for developers fine-tuning and deploying open models, regardless of which company’s headline number was used.