Google DeepMind’s Gemma family of open AI models has passed one billion cumulative downloads, the company said this week, marking two years since the lightweight, freely downloadable models first launched as an alternative to closed systems like Google’s own Gemini.
A Milestone Two Years in the Making
Google DeepMind vice president Clement Farabet and product director Olivier Lacombe disclosed the figure in a blog post, saying developers have published more than 100,000 distinct Gemma variants since the family launched in February 2024. Google described the ecosystem of derivatives, fine-tunes and adaptations as the “Gemmaverse.” Unlike Google’s Gemini, which runs on the company’s servers, Gemma is released with open weights, meaning anyone can download, inspect and run the models on their own hardware.
Growth has accelerated sharply in recent months. Gemma reached 150 million downloads by May 2025 and 500 million by April 2026, according to earlier company disclosures. Gemma 4, launched in April 2026, has alone accounted for more than 300 million of the total downloads since its release, meaning a single generation of the model family drove more than a third of its entire lifetime total.
From Space to the Ocean Floor
Google highlighted a range of applications far removed from typical chatbot use. Teams at NASA’s Jet Propulsion Laboratory, satellite startup Satlyt and orbital computing company Starcloud are running Gemma models directly aboard satellites, using them for onboard image analysis, deciding which images are worth transmitting back to Earth over limited bandwidth, and routing communications between spacecraft. NASA flew a compressed version of Gemma 3 on a Loft Orbital satellite earlier this year in what was reported as the first in-orbit demonstration of a vision-language model analysing imagery from a satellite’s own camera.
Other uses cited by Google include DolphinGemma, a model developed with Georgia Tech and the Wild Dolphin Project to analyse dolphin vocalisations, and C2S-Scale, a Yale collaboration that used Gemma to help identify a novel cancer therapy pathway later verified in living cells. In India, the National Health Authority has built Gemma 4 into Aarogya Setu 2.0, a health records app with more than 100 million downloads on Android.
Still Behind the Leader
Despite the milestone, Gemma remains well behind Alibaba’s Qwen family in raw download numbers. Alibaba said in mid-August that Qwen had surpassed 3 billion downloads, with more than 300,000 derivative models published by outside developers, a figure that independent tracking from Hugging Face put closer to 2 billion downloads for Qwen this year alone. Google did not break down its own billion-download figure by model version, year or platform, and a Google engineer noted on social media that the count excludes Gemma integrations built into Android and Chrome.
To mark the milestone, Google launched a new “Awesome Gemma” directory on GitHub intended to catalogue community projects, fine-tunes and developer tools. The company also said its most recent Gemma Challenge competition on the Kaggle platform drew more than 1,600 project submissions, including tools such as AI assistants for visually impaired users and offline educational resources for areas with limited internet connectivity.
Why It Matters
Download counts are an imperfect measure of real-world usage, since model weights are often mirrored and re-uploaded across multiple platforms, and a single download says little about how often a model is actually run. Even so, the billion-download figure underscores how central open-weight models have become to Google’s broader AI strategy, giving the company a foothold among developers who want to run models locally rather than through a paid API, even as it continues to compete for enterprise customers with its closed Gemini models.
