Gemma's Open Models Cross a Billion Downloads

Google's family of open models has passed a milestone that even its own researchers seem a little surprised by. On August 20, the company said Gemma had crossed one billion downloads in roughly two years, with outside developers publishing more than 100,000 variants and fine-tunes built on the open weights.

The number is cumulative rather than a count of active users, so it flatters the story a little. Still, it puts Gemma in rare company. Google now runs two very different billion-scale franchises at once: the closed Gemini assistant, which recently passed a billion monthly users, and Gemma, the weights anyone can download, run locally and rebuild.

What people are actually doing with it

The announcement leaned on examples rather than benchmarks, which is telling. Google DeepMind vice president Clement Farabet and product director Olivier Lacombe pointed to teams at NASA, Satlyt and Starcloud running Gemma directly on satellites, where it handles onboard image analysis and helps manage scarce downlink bandwidth. India's National Health Authority folded Gemma 4 into Aarogya Setu 2.0, a health app with more than 100 million Android installs, using it to turn messy medical reports into standardized digital records.

The research example is the one worth pausing on. Scientists at Yale and Google used a Gemma-based model called C2S-Scale to read the "language" of single cells, and it flagged a novel cancer therapy pathway that later held up when tested in living cells. That is the kind of result open weights are supposed to enable: a specialist team taking a general model somewhere its makers never planned to go.

The ecosystem is the point

Google has taken to calling this sprawl of derivatives the "Gemmaverse," and it launched an "Awesome Gemma" repository on GitHub to help people navigate it. The branding is a bit much, but the strategy underneath is serious. Open weights have become a competitive battleground, with Chinese labs shipping capable systems like Z.ai's GLM-5.3 and Meta putting an open agent model on laptops. Downloads are how a lab keeps its architecture in front of the developers who will build the next wave of products.

There is a tension here that Google mostly left unspoken. The same openness that let a cancer-research team push Gemma forward is the openness critics have in mind when they ask whether open models hand rivals a weapon. A billion downloads is a vote of confidence from developers. It is also a reminder that once weights are out, nobody controls where they go.