Google’s Gemma Open Models Pass 1 Billion Downloads as Variants Top 100K

Google DeepMind’s Gemma family of open models has passed one billion cumulative downloads, the company said on August 20, 2026, capping two years in which outside developers published more than 100,000 variants built on the models’ open weights. The figures, disclosed in Google’s official announcement by Clement Farabet, a vice president at Google DeepMind, and product director Olivier Lacombe, mark the first time Google has put a cumulative total on Gemma adoption since the family launched in early 2024.

Download counts are an imperfect measure of what an open model actually does — weights get pulled, mirrored, and re-uploaded across hubs, and one download says nothing about how often a model runs. The number Google paired with the milestone is the more telling one for the ecosystem’s shape: over 100,000 distinct Gemma variants published by developers, fine-tunes and derivatives adapted to specific languages, tasks, and hardware targets. That variant count is the difference between a model people try and a model people build on.

The post is structured around where those variants are running, and the geographic spread is unusual. Teams at NASA, satellite startup Satlyt, and orbital-compute company Starcloud are running Gemma models in space: for onboard image analysis, for deciding what is worth sending down scarce downlink bandwidth, and for routing communications between satellites. The NASA case is the best documented of the three. The agency’s Jet Propulsion Laboratory flew a 4-bit compressed version of Gemma 3 4B on a Loft Orbital satellite earlier in 2026, in what IEEE Spectrum reported on July 23, 2026 as the first in-orbit demonstration of a vision-language model analyzing imagery from a satellite’s own sensor. The system, called NAVI-Orbital, classified images with 88 percent accuracy in a ground benchmark of 7,960 images and ran live captures over Toulouse, France, and the coast of Argentina on an Nvidia Jetson Orin AGX module, hardware constrained enough that the 4-billion-parameter model’s 8-gigabyte memory footprint was the enabling specification.

On the ground, the largest single deployment the announcement cites is in India, where the National Health Authority integrated Gemma 4 and Google’s open Medical Data Toolkit into Aarogya Setu 2.0, an app with more than 100 million Android downloads, to convert medical reports into standardized digital formats that patients can share across providers. In research, Yale and Google researchers built C2S-Scale, a model trained to interpret single-cell data, on top of Gemma; Google says it produced a cancer therapy pathway later verified in living cells. The company’s domain-specific MedGemma models, meanwhile, are being used in clinical application development ranging from outpatient triage at the All India Institute of Medical Sciences to tools for frontline health workers in rural Uganda. A fourth project, DolphinGemma, built with Georgia Tech and the Wild Dolphin Project, applies a Gemma variant to predicting sequences in dolphin vocalizations (a long-running research effort rather than a shipped product).

A Milestone That Doubles as a Product Directory Launch

Alongside the numbers, Google used the post to launch the Awesome Gemma repository on GitHub, a curated directory of community projects, fine-tunes, tutorials, and tools that it is positioning as the official index of what it calls the Gemmaverse. The company also noted that its recent Gemma Challenge on Kaggle drew more than 1,600 project submissions, with winners to be announced in the coming weeks.

The framing matters for how to read the milestone. Google does not sell Gemma; the family is the company’s answer to Meta’s Llama in the contest to be the default substrate for open-weights development, and download counts are the scoreboard both labs publish.

What the billion-download figure does not capture is how much of that volume translates into sustained use, and Google offered no breakdown by model generation or platform. The more durable signal is in the deployments the company chose to feature: the same weights running on a radiation-hardened satellite computer drawing a few hundred watts, inside a national health app, and in a cell-biology research pipeline. That range — from 4-billion-parameter edge models to research-scale systems — is the argument Google is making for why open weights remain strategically central to its model line even as its frontier Gemini systems stay closed. With Gemma 4 now the current generation and a centralized community directory live, the next observable markers are the Kaggle challenge results and whatever adoption figures Google attaches to Gemma 4 specifically in the months ahead.