Google's open Gemma model family has surpassed one billion downloads. Documented uses range from phones to embedded infrastructure, and even space applications, as well as research into animal communication and medical work. A community repository has been opened to gather these variants.
We have covered the rise of open models from the angle of releases and performance. Here is the adoption angle, and it tells a different story.
What this figure measures, and what it doesn't
A download is not a user. The same developer can download ten variants, an automated system can fetch a model at every deployment, and many downloads lead nowhere.
What this figure really measures is the volume of experimentation. And at this scale, it indicates that these models have entered ordinary software engineering practice rather than remaining demonstration objects.
This is consistent with what we reported about half of the Fortune 500 favouring open models. The download figure is the technical trace of that shift.
What the diversity of uses reveals
The list of applications is more instructive than the total. Phones, embedded hardware, scientific research, space applications.
None of these uses would have been possible with a model accessible only through a remote service. An isolated sensor, an onboard instrument, a device without a reliable connection: these are contexts where the API does not exist as an option.
This is the argument we developed about local AI and pooling machines: open weights are not just about saving money, they make possible situations where the cloud makes no sense.
This figure should also be read for what it is: a strategic success. A company that publishes open models while its flagship keeps slipping gains through this channel what it struggles to gain through the other: a presence in projects, developer familiarity, a default position. Publishing openly is not a commercial renunciation, it is a different way of occupying the ground. We noted this regarding Ox Alpha buying its way into workflows.
What this changes in the debate
This milestone brings a concrete element to a discussion that is often theoretical.
Proponents of closed models point to irreversibility: a published model cannot be recalled. That is true, and we have documented it with an open model integrated into an attack tool.
This figure shows the other side of the scale: one billion downloads, almost all for mundane and useful purposes. The debate over publishing weights is usually conducted from the extreme cases. The bulk is elsewhere, and it is ordinary.
What to take away
This milestone confirms a shift we have been describing for months: open models are no longer a militant alternative, they are a standard engineering building block.
It also reminds us where real democratisation is happening. Not in the trillion-parameter models whose releases we cover, but in those that fit on a phone and that no one notices. The progress that matters to the greatest number is almost always the one that doesn't make headlines.