Google's open-weights Gemma family passed one billion cumulative downloads on Wednesday, August 20, 2026, the company disclosed in a blog post by Google DeepMind vice president Clement Farabet and product director Olivier Lacombe. It is the first time Google has put a headline number on Gemma adoption since the family launched in early 2024, and it lands as open-model releases from Meta, Alibaba's Qwen team and Alibaba's HappyShrimp are competing for the same developer mindshare.
The Gemmaverse: 100,000-plus community variants
Alongside the download milestone, Google said developers have published more than 100,000 distinct Gemma variants — fine-tunes and derivatives tuned for specific languages, tasks and hardware targets — which the company now brands the “Gemmaverse.” To make that ecosystem easier to navigate, DeepMind launched an Awesome Gemma repository on GitHub as an official directory of community projects, tutorials and developer tools. Gemma hit 150 million downloads in May 2025 and 500 million by April 2026, meaning the family has doubled adoption in roughly four months.
From orbit to India's public health stack
Google's post details several flagship deployments that highlight where the on-device economics of Gemma pay off. NASA, satellite imaging start-up Satlyt and orbital-compute company Starcloud are running Gemma directly on satellites for onboard image triage and inter-satellite routing, cutting scarce downlink bandwidth. India's National Health Authority has integrated Gemma 4 and Google's open-source Medical Data Toolkit into Aarogya Setu 2.0, an Android app with more than 100 million installs, to standardise medical reports for citizens. And a Yale–Google collaboration used a Gemma-based model called C2S-Scale to identify a novel cancer therapy pathway later verified in living cells — described as the first AI-derived mechanistic pathway confirmed at the bench.
Domain-specific spin-offs and dolphin talk
DeepMind's medical variant MedGemma is now in use at the All India Institute of Medical Sciences for outpatient triage and by frontline health workers in rural Uganda via the CRANE-AI project. On the more exotic end, DolphinGemma — built with Georgia Tech and the Wild Dolphin Project — is being trained to predict sequences of dolphin vocalisations. The Gemma 4 release earlier this year added stronger reasoning and agentic behaviours, positioning the family as Google's answer to Anthropic and Meta's Llama series in the open-weights race.
Reporting based on coverage from The Keyword (blog.google), unite.ai and Stocktwits.