Google says its Gemma family of open models has passed a billion downloads.

Developers have published more than 100,000 Gemma variants over the past two years, the company said in a blog post on Thursday. Clement Farabet, a vice president at Google DeepMind, and Olivier Lacombe, a product director there, wrote it.

Google calls the ecosystem around those variants the Gemmaverse.

“What matters far more than the download count is what the community is building with them,” the two wrote.

What the number counts

Gus Martins, a Google developer relations engineer, added a caveat on Bluesky.

“1B Gemma downloads,” he wrote. “Fun fact: this doesn’t count Android and Chrome integrations.”

Google did not break the total down by model, by year, or by platform.

Omar Sanseviero, who works on Gemma at Google DeepMind, described the spread on X. “From running underwater to running at space,” he wrote.

How that compares to Qwen

Alibaba made a larger claim five days earlier.

TNW reported on 16 August that Alibaba said Qwen had passed 3 billion downloads. Hugging Face’s State of Open Models report counted 2,045 million Qwen downloads this year.

On derivatives, Alibaba claimed more than 300,000. Hugging Face counted 151,448 Qwen-based derivatives on its hub, which the report said was 2.6 times Meta’s entire footprint there.

Google’s figure of 100,000 Gemma variants covers two years and is the company’s own count.

Hugging Face says its download figures measure activity inside its own hub. They do not capture API usage, private deployments or models distributed elsewhere.

Gemma in orbit

Google says NASA, Satlyt and Starcloud are running Gemma directly in orbit.

The uses it lists are onboard image analysis, optimising scarce downlink bandwidth and routing communications between satellites. Satlyt and Starcloud are both startups; NASA is the agency.

Google says those deployments show Gemma can handle complex reasoning in constrained environments.

It does not say which Gemma models are running on which spacecraft, or since when.

Gemma in India’s health app

India’s National Health Authority has integrated Gemma 4 and Google’s open-source Medical Data Toolkit into Aarogya Setu 2.0.

Google says that app has over 100 million downloads on Android. Gemma 4 turns medical reports into standardised digital formats, which Google says lets citizens manage and share health records across providers.

The cancer pathway claim

Researchers from Yale and Google built a model called C2S-Scale on Gemma, designed to interpret the language of single cells.

Google says C2S-Scale “successfully discovered a novel cancer therapy pathway that was verified in living cells”. It calls that the first time an AI system produced novel mechanistic therapeutic pathways verified in living cells.

The claim appears in Google’s own announcement. In this post the company links to its earlier write-up of the work rather than to a paper.

MedGemma and DolphinGemma

MedGemma is Google’s domain-specific medical model.

Google says organisations are using it in clinical applications. The examples it gives are outpatient triage at the All India Institute of Medical Sciences, and systems supporting frontline health workers in rural Uganda.

DolphinGemma came out of a collaboration with Georgia Tech and the Wild Dolphin Project. Google says the model processes dolphin vocalisations to predict sound sequences. It describes the work as ongoing and says researchers are still iterating.

The Awesome Gemma repository

Google has launched a GitHub repository called Awesome Gemma as the official directory for the Gemmaverse. It says the repository will feature community projects, fine-tunes, tutorials and developer tools.

Philipp Schmid, who works on Google’s AI developer experience, listed the contents on X. Model cards and collections for 16 Gemma variants, setup guides for Ollama, vLLM and LiteRT, and fine-tuning recipes for Unsloth, Tunix and MLX.

TNW reported in July that Ollama raised $65mn as its open-model runner reached nearly 9 million developers.

Google also says its Gemma Challenge on Kaggle drew more than 1,600 project submissions, with winners to be announced.

Ollama posted its own note. “Gemma is one of the most popular open models,” the company wrote. “It’s been an amazing journey being a close partner.”

Where Gemma runs

Google says Gemma was built to deploy across environments, from local devices to edge infrastructure.

TNW reported this month that Google built an offline translator running on a Raspberry Pi using Gemma.

Lacombe described the original goal on X. “Build open models that are lightweight, highly capable, and flexible enough to run anywhere,” he wrote.

In last year’s equivalent post, Google highlighted developers building AI assistants for visually impaired people and offline educational hubs for regions without connectivity.

The open-model field around it

Other labs have been shipping open models through the same period.

Nvidia is working towards a trillion-parameter open model, TNW reported this month. Nvidia has said it is chasing China on open weights.

The policy position moved this week too. The White House technology strategy published on 20 August left open-weight AI off its critical list.

What Google has not said

Google has not published the methodology behind the billion figure. It has not named the platforms the count draws from, and Martins’s note rules out at least two Google surfaces.

It has not said how the 100,000 variants were counted, or how many are actively used.

Google has not given usage figures for the deployments it names. The one number it does give, 100 million Android downloads, belongs to Aarogya Setu 2.0 rather than to Gemma.

Google’s own announcement is the source for the C2S-Scale result, the orbital deployments and the clinical uses. TNW has not independently verified any of them.

Get the TNW newsletter

Get the most important tech news in your inbox each week.