Related: A Chinese Open Model Now Matches Anthropic’s Mythos at Finding Security Bugs
Google’s open AI model Gemma 4 is being adopted at a striking rate, with the company saying it has crossed 200 million downloads in only about two and a half months since launch. For a model family released in early April, that pace places it among the fastest spreading open models yet, and it offers a clear signal of how central freely downloadable AI has become for the developers and companies building on top of it. The number is Google’s own, but the trajectory it describes fits a broader shift toward open weights that has been gathering force all year.
Gemma 4 arrived on April 2 under the permissive Apache 2.0 license, which lets anyone download, run, and modify the models without paying or asking permission. Google DeepMind positioned the family as its most capable open release to date, built to be efficient enough to run on ordinary hardware rather than only in the cloud. The models are multimodal, able to take in images and video across the lineup, with native audio understanding on the smaller variants designed for phones and laptops, a combination meant to let developers build richer applications without reaching for a closed, paid system.
The appeal of an open model is as much practical as ideological. A developer who downloads Gemma 4 can run it on their own machines, keep their data in house, fine tune it for a narrow task, and ship it inside a product without per query fees or rate limits. That control matters most to the companies and researchers who cannot or will not send sensitive data to someone else’s servers, and it is a large part of why open models have carved out a durable place alongside the frontier systems from Google and its rivals. Two hundred million downloads is less a measure of how good the model is than of how many people found it worth building on.
The momentum extends well beyond this one release. Across its generations, the broader Gemma line has now surpassed 400 million total downloads and spawned more than 100,000 community built variants, a sprawling ecosystem that Google has taken to calling the Gemmaverse. Each fine tuned variant is a sign that someone took the base model and adapted it to a specific job, from a niche language to a specialized coding assistant, and that long tail of customization is exactly what open releases are designed to encourage. The base model becomes a starting point rather than a finished product.
The reach has been helped by where Gemma 4 can be found. Beyond a direct download, the models have been made available through major cloud platforms, including a listing on Amazon’s Bedrock service in mid June, and through Google’s own on device frameworks for Android. Meeting developers where they already work lowers the friction of trying a model, and it helps explain how a download count climbs into the hundreds of millions so quickly. An open model that is easy to obtain in every environment a developer might use will spread faster than one that demands a special setup.
A download is not the same as active use, and the figure should be read with that in mind. Many downloads are experiments, duplicates pulled into automated pipelines, or models that get tried once and set aside, so the raw count overstates how many real applications are running on Gemma 4 at any moment. Google also has every incentive to highlight a large number, since open models are part of how it competes for the loyalty of developers who might otherwise standardize on a rival’s tools.
Even with that caveat, the milestone says something real about the state of the field. The choice between open and closed AI has become one of the defining questions of the year, and a free model racking up 200 million downloads in a quarter shows that the open path is not a niche preference but a mainstream one. For Google, keeping developers building inside its ecosystem, even with a free model, is a strategic prize, and Gemma 4’s fast start suggests the bet is paying off.
Related on Entrelligence: Google’s Nano Banana 2 Lite and Omni Flash models, a Chinese open model matching Anthropic’s Mythos, and Cursor’s iOS app for AI coding agents.
Your First 10 AI Skills
10 practical AI skills, copy-paste prompts and a 7-day plan to start using AI with confidence.
Download the guide →