Google's Free Models Passed a Billion Downloads. Some of Them Are Running in Orbit
Gemma crossed one billion downloads and 100,000 community variants in two years. Alibaba's Qwen claimed three billion five days earlier.
Google DeepMind announced on 20 August that its Gemma family of open models has passed one billion cumulative downloads. In the two years since the first release, developers outside Google have published more than 100,000 variants built on those weights, an ecosystem the company calls the Gemmaverse.
Open models are the ones you can download and run yourself. Instead of sending every request to a company’s servers, you get the weights, the numbers that make up the trained model, and run them on your own machine, your own server, or a device with no internet connection at all. Gemma models are small enough to make that practical, which is why the deployment examples in Google’s post are so unusual. NASA, the satellite startup Satlyt and the orbital compute company Starcloud are running Gemma on satellites, using it to analyse images on board and decide what is worth sending down over scarce bandwidth. India’s National Health Authority built Gemma 4 into Aarogya Setu 2.0, an app with more than 100 million Android downloads, to turn messy medical reports into a standard digital format. Researchers at Yale and Google used a Gemma-based model called C2S-Scale to identify a cancer therapy pathway that was then verified in living cells.
Google also launched an “Awesome Gemma” repository on GitHub as an official directory of community projects, fine-tunes and tutorials.
Download counts are a soft metric and everyone knows it. One download is not one user, automated pipelines pull the same weights repeatedly, and Google notes that the figure excludes Android and Chrome integrations, which cuts the other way. Read it as a rough measure of ecosystem gravity, nothing more. The comparison that puts it in proportion arrived five days earlier: Alibaba said its Qwen family had passed 3 billion downloads with over 300,000 derivatives, roughly triple Google’s numbers. Both companies are competing for the same thing, which is not revenue from the models themselves but developers who build on their stack, learn their tooling and eventually pay for something adjacent. Open weights are the customer acquisition strategy.
What this means for you: if you have never tried running a model on your own hardware, Gemma is one of the gentler places to start, and the smaller sizes run on an ordinary laptop. The real appeal is not cost but control: nothing leaves your machine, which matters for medical notes, client documents, or anything you would rather not upload. A fair caveat, since the milestone invites hype: an open model you run yourself will be noticeably less capable than the best hosted models, and setting it up takes an afternoon rather than a click. For most everyday use the hosted assistants remain easier. But for the specific case of “this data must not leave the building”, the free option is now genuinely good enough.
Sources
Source: https://blog.google/innovation-and-ai/technology/developers-tools/gemma-one-billion-downloads/
Researchers Encrypted the Attack So Grok Could Not Read It. Grok Decrypted It Itself
Adversa AI hid instructions in a webpage using AES-256. Grok's own code sandbox unlocked them and then leaked the user's chat history. Reported in June, still unpatched.
