Tech and AI
An open model family passes a billion downloads, and the places it is running now include an orbiting spacecraft and an Indian health app used by 100 million people
By Staff Writer | 22 August 2026
Google DeepMind said on 20 August that its Gemma family of open models has gone past a billion cumulative downloads, and that developers have published more than one hundred thousand variants of them over two years. The two executives who signed the announcement put the weight on what is being built rather than on the count. The named deployments run from onboard image analysis in orbit to a single cell model that produced a cancer therapy pathway later verified in living cells.
Open weights are models a developer can download and run on their own hardware, and the argument for publishing them has always been that the interesting uses are the ones the publisher did not think of. This announcement is that argument set out with names attached, which is rarer than the argument itself.
What matters far more than the download count is what the community is building with them.
Clement Farabet, Vice President, and Olivier Lacombe, Product Director, Google DeepMind
Running in orbit, and in a national health record
The announcement states that teams at NASA, at Satlyt and at Starcloud are running these models directly in orbit, doing onboard image analysis, working out how to use scarce downlink bandwidth and routing communications between satellites. The claim being made is about the operating conditions rather than about the science: a model small enough and reliable enough to reason on a spacecraft is a model that will run on almost anything on the ground.
The second deployment is a good deal closer to ordinary life. India's National Health Authority has put Gemma 4 and an open source medical data toolkit into Aarogya Setu 2.0, an app the announcement describes as having more than 100 million downloads on Android. The job the model does there is unglamorous and useful: it turns complicated medical reports into standardised digital records so that a citizen can hold their own data and pass it between providers.
A therapy pathway that was checked in living cells
The research claim is the one worth reading carefully. Researchers at Yale and at Google built a model called C2S-Scale on these weights to interpret single cell data. The announcement says it discovered a novel cancer therapy pathway that was verified in living cells, and that this was the first time an AI system produced novel mechanistic therapeutic pathways verified in living cells. That is a claim about a first, made by the party with an interest in it, and it is reported here as their claim rather than as a settled fact.
A domain model called MedGemma is described as in use by organisations building clinical applications, with two examples named: supporting outpatient assessment at the All India Institute of Medical Science, and supporting frontline health workers in rural Uganda. A further collaboration with Georgia Tech and the Wild Dolphin Project produced DolphinGemma, which processes dolphin vocalisations to predict sound sequences, and which the announcement describes as work still being iterated rather than finished.
The part that is a directory, and why it matters
Alongside the milestone the company is opening a curated repository on GitHub, described as the official directory for the community's projects, fine tunes, tutorials and developer tools. It also reports more than 1,600 submissions to a recent community challenge, with winners still to be announced.
For anyone assessing whether to build on open weights, the useful signal in all this is not the billion. It is that four of the named deployments are in regulated or safety sensitive settings, which means somebody has already had the argument about assurance, about what the model is allowed to decide and about what happens when it is wrong. The announcement does not describe those controls, and a professional adviser should assume nothing about them from the fact that the deployment exists.