Gemma passes a billion downloads and 100,000 derived models

On August 20, Google announced that the Gemma family of open models passed a billion downloads. The number itself is a celebration, but what sits behind it points to a change that reaches people who build sites: a model running on someone else's server stops being the only option.

The figures behind the announcement

• Models in the Gemma family have been downloaded a billion times.
• In two years the community has published more than 100,000 derived models.
• The most recent competition on Kaggle drew more than 1,600 entries.
• An Indian health application that embedded Gemma has been downloaded a hundred million times.

Where it is actually used

The announcement lists several uses outside the usual frame. NASA, Satlyt, and Starcloud run Gemma in orbit, processing satellite imagery on the spot. India's National Health Authority embedded Gemma 4 in its health platform. Researchers at Yale used C2S-Scale, derived from Gemma, in work on pathways in cancer treatment, and the DolphinGemma project works on understanding dolphin vocalization. MedGemma, meanwhile, is aimed at clinical use, from triage to supporting health workers in the field.

Awesome Gemma

An Awesome Gemma repository launched with the announcement. It is meant as a curated index of community projects, fine-tuned models, tutorials, and tools — the place you start when checking whether someone has already solved the same job. For a family with more than a hundred thousand derived models, an index like that is as useful as the models themselves.

Why it matters for site work

When an open model can run on your own server, jobs like routing form submissions, summarizing comments, or detecting language stop requiring that content go to a third party. That removes an entire section from the personal data conversation with a client. Cost is no longer per call but per rented server, which often works out better for steady, predictable work.