A billion downloads, and nobody will say what a download is

Two vendor totals and one independent count are on the record, and only one of them says what it measured. Hugging Face's number is activity inside its own hub, which leaves out API usage, private deployments and models distributed anywhere else ^[11]. That makes it narrow, and it makes it checkable. The vendor totals are neither.

The arithmetic is where this gets awkward. Alibaba's 3 billion is roughly 1.5 times the 2,045 million Qwen downloads Hugging Face logged this year, a gap of about 955 million ^[13]. On derivatives the multiple is worse: the more than 300,000 Alibaba claims is close to twice the 151,448 Qwen-based derivatives Hugging Face found on its hub ^[9][14]. Both pairs of figures are presented as descriptions of the same ecosystem, and neither publisher explains the difference.

Gus Martins of Google's developer relations team supplied the most useful sentence in the Gemma announcement, on Bluesky: the billion does not count Android and Chrome integrations ^[5]. Read it twice. It asserts the total is an undercount, which is entirely plausible, and it concedes that the boundaries of the count are known only inside Google. The company published no split by model, by year or by platform ^[6]. There is a real off-hub channel here too: Ollama's open-model runner had reached nearly 9 million developers when it raised $65mn in July, according to TNW ^[16], and a pull through a local runner is not a line in anyone's hub statistics.

For the same reason, the temptation to rank Gemma against Qwen should be resisted, even though the numbers invite it. Gemma's claimed cumulative billion ^[1] is about half the Qwen total Hugging Face counted in a single year on a single hub ^[15]. Those denominators are not the same thing, and the comparison collapses if you press on it.

The pattern in the announcement is self-certification. The 100,000-plus Gemma variants figure is Google's own count over two years ^[12]. Google says C2S-Scale, built on Gemma with Yale researchers, discovered a novel cancer therapy pathway verified in living cells, and calls it a first ^[18]; the post links to the company's earlier write-up rather than to a paper ^[19]. Google says NASA, Satlyt and Starcloud run Gemma in orbit, without saying which models on which spacecraft, or since when ^[21].

What is checkable sits in the smaller numbers. The new Awesome Gemma repository, per Google's Philipp Schmid, carries model cards for 16 Gemma variants, setup guides for Ollama, vLLM and LiteRT, and fine-tuning recipes for Unsloth, Tunix and MLX ^[17]. An engineer can verify that in an afternoon and act on it. The Kaggle Gemma Challenge drew more than 1,600 submissions, with winners still to come ^[22]. Even the strongest deployment claim is denominated in downloads: India's National Health Authority has put Gemma 4 and Google's Medical Data Toolkit into Aarogya Setu 2.0, an app Google says has more than 100 million Android downloads ^[20].

Clement Farabet and Olivier Lacombe wrote that what matters far more than the download count is what the community is building with the models ^[3][4]. The download count is the headline anyway.