A nearby count, not verification of the 52%
The figure moving through aggregators today says Alibaba's models took 52% of downloads "among the most popular chat models" on Hugging Face over the past month, more than seven times the share given to Google and to OpenAI, at 7% each. Read quickly, it sounds like half of everything downloaded from the Hub is Alibaba's. It is not. The circulating claim's own calculation could not be traced. The nearest count that can be reproduced rests on a defined table, a defined window and a defined way of crediting a download to a family.
The most transparent version is VIDRAFT's Global LLM Download Leaderboard, which ranks Hugging Face text-generation repositories by trailing 30-day downloads and assigns each repository to a model family using its architecture-based classification. On November 4, the Qwen family held 50.9% of downloads across the Top 300. That is genuinely close to 52%. It is also a share of 300 repositories, not of the whole Hub, and it counts community quantizations and repacks toward Qwen, because they carry Qwen's architecture.
The company's own share is a third, not a half
Split the same table differently and the story changes. Counting only repositories uploaded by the Qwen organization itself, the share on November 4 was 34.5%. The gap between 50.9% and 34.5% is third-party repositories assigned to Qwen: other uploaders' fine-tunes, quantizations and repackaged builds of Qwen weights. That gap is arguably the more interesting fact, because it shows Qwen's position extends well beyond Alibaba's own uploads. But it also means "Alibaba took 52%" would credit the company with downloads of files it did not upload, measured on a table it did not define.
One small old model moves everyone's share
The table has a distortion worth knowing about before quoting any of its percentages. The single most-downloaded row on November 4 was openai-community/gpt2, a 0.14B-parameter model, with 29.8 million downloads in 30 days and 8.7% of the Top-300 total. VIDRAFT's own reading is that this is mostly automated traffic: tutorials, test suites and libraries fetching a small, always-available model. Hugging Face's counter cannot separate that traffic from a person choosing a model, and neither can any chart built on it.
The distortion is large enough to flip conclusions. With gpt2 included, the GPT/OPT family appears to gain share, rising to 15.4%. Remove that one repository and the same family falls to 7.3%. The GPT/OPT category is not equivalent to OpenAI's repositories or business, so its percentage cannot validate the circulating company comparison. Qwen's recent slide also halves, from a 4.9-point drop to a 2.2-point drop, once gpt2 is excluded.
The direction is still down, slightly
Two snapshots 48 days apart cannot separate a trend from a bump, and VIDRAFT says so itself. Within that limit, Qwen's share fell from 55.8% to 50.9%, and its absolute downloads fell 21%, faster than the table's overall 13.6% decline. Gemma lost the most share, from 5.6% to 2.5%. DeepSeek rose from 4.6% to 6.7%. The trackers report broadly similar share levels under different methods; they do not establish a common trend. The honest sentence is "about half of a defined table," not "52% and dominant."
A second, wider tracker reports a similar level under a different method. Azumo's tracking of the top 5,000 models put Qwen-based models at 45 to 48% of open LLM downloads in each month from March to September 2026, and at 47.2% in the week of October 26. Different denominator, similar answer: high forties to low fifties, depending on where you draw the table's edge.
What a download is not
None of these numbers count people. These are repository download counts, which can include CI pipelines, mirrors and automated tests. They also exclude API usage and distribution outside Hugging Face entirely, so they cannot rank companies, only tables.
Sources
- [1] Strong Tower AI, AI Company News (Nov 15, 2026), summarizing AI Weekly / Yahoo Finance dataRead source
- [2] VIDRAFT, "Qwen Still Takes Half of Open-LLM Downloads: What Changed in the Hugging Face Top 300 Since September" (snapshots Sep 17 and Nov 4, 2026)Read source
- [3] Azumo, "Most Downloaded AI Models on Hugging Face" (top-5,000 tracking, updated Nov 2026)Read source
- [4] Metaprompt, "Chinese Open-Weight Models: Downloads, Derivatives and Token Share" (method comparison, Hugging Face counts retrieved Oct 13, 2026)Read source