Google DeepMind said Thursday that its Gemma family of open models has passed one billion downloads, and that developers have published more than 100,000 variants built on those open weights over two years. The post came from Clement Farabet, a vice president at DeepMind, and Olivier Lacombe, a product director there. Google calls the ecosystem around it the Gemmaverse.
Five days earlier, Alibaba said Qwen had passed three billion.
That is the whole story in two lines. The open-model race now has a public scoreboard, both sides are keeping their own score, and the American entry is announcing a number the Chinese entry beat by a factor of three the week before.
The numbers nobody audited
Google has not published the methodology behind the billion. It has not named the platforms the count draws from, has not broken it down by model or by year, and has not said how the 100,000 variants were counted or how many are actively used. Gus Martins, a Google developer relations engineer, added on Bluesky that the total does not include Android and Chrome integrations — which cuts the other way, and rules out two of Google’s own largest surfaces.
Alibaba’s three billion carries the same problem. Both are self-reported.
The one place there is an outside count is Hugging Face, whose State of Open Models report put Qwen at 2,045 million downloads this year and 151,448 Qwen-based derivatives sitting on its hub — a derivative footprint the report described as 2.6 times Meta’s entire presence there, against Alibaba’s own claim of more than 300,000. Hugging Face is explicit that its figures measure activity inside its own hub. They do not capture API calls, private deployments, or models distributed anywhere else.
Our take: Download counts are the open-model equivalent of app-store installs — easy to inflate, impossible to falsify, and only loosely connected to production usage. The number that actually matters is the one AT&T published this week: 40% of 45 billion daily tokens routed to open weights, at a 56% cost cut and a 2% quality drop. That is a measured workload, not a marketing metric. Treat every billion-download claim as an ecosystem-health signal and nothing more.
What Google is actually pointing at
Read past the headline number and the announcement is a deployment list. Google says NASA and the startups Satlyt and Starcloud are running Gemma in orbit for onboard image analysis, downlink-bandwidth optimisation and intersatellite routing. India’s National Health Authority has integrated Gemma 4 and Google’s open-source Medical Data Toolkit into Aarogya Setu 2.0. Yale and Google researchers built a model called C2S-Scale on Gemma, which Google says found a novel cancer therapy pathway later verified in living cells.
Those are Google’s claims, sourced to Google’s own post. It has not said which Gemma models run on which spacecraft, or since when, and the cancer result links to an earlier company write-up rather than a paper. The company also shipped an “Awesome Gemma” GitHub repository as the official directory for the ecosystem, and says its Kaggle Gemma Challenge drew more than 1,600 submissions.
The strategic read is simpler than the deployment list. Small open models that run on a laptop, an edge box or a satellite are the distribution layer for everything above them, and Chinese open weights have been winning that layer on price. Publishing a billion-download post five days after a three-billion-download post is not a coincidence.
What to watch
- Whether Google or Alibaba ever publishes methodology. Neither has, and the gap between self-reported totals and Hugging Face’s hub counts is where the real number lives.
- Derivative counts over download counts. A fine-tune is evidence somebody built something; a download is evidence somebody clicked.
- Washington’s posture. The White House technology strategy published 20 August left open-weight AI off its critical-technology list — a quiet signal after the industry letter that 25 companies signed and the three biggest closed labs did not.
- Enterprise routing data. Token-share numbers from companies actually paying the bill will settle this argument long before download counters do.