Google has released Gemma 4 today, April 2, and the announcement contains something more useful than another invitation to admire a chatbot: downloadable models that developers can run and adapt themselves.
The headline numbers are impressive, but one word matters. In Google’s launch announcement, the 31B version places third among open models on Arena’s text leaderboard, with the 26B version sixth. That is not a claim that they rank third and sixth across every open and proprietary system.
A leaderboard can be useful without becoming a royal coronation. Your actual task may involve a difficult spreadsheet, a particular language or a codebase that refuses to resemble the benchmark.
Four Models, Different Jobs
The launch family consists of E2B, E4B, 26B Mixture of Experts and 31B Dense. Google presents the smaller pair as options for on-device work, while the larger models target more demanding setups. The 26B model activates about 3.8 billion parameters during inference.
That last figure describes active computation, not permission to forget the rest of the model when budgeting memory. Likewise, the “E” in the smaller names means effective: it should not be read as a complete count of everything you must store.
All four accept images and video; the smaller two also accept audio. Google lists context windows of 128K for the edge models and up to 256K for the larger pair. A large context window is room for input, not a promise that every repository fits or that every detail will be handled correctly.
Local AI Still Needs a Suitable Home
Choosing a model involves more than finding the biggest number your download folder can accommodate. Available memory, the inference software, the precision of the weights and the length of your prompts all affect the experience.
Start with a modest task you can judge: summarize a document you know, explain a small piece of code, or extract information from a clear image. Increase the workload after seeing how your setup behaves. A laptop fan auditioning for a jet engine is useful feedback, even if it is not a formal benchmark.
It is another example of the growing interest in open AI models. Downloadable weights give you choices; they do not make every model equally comfortable on every computer.
Running it yourself is also different from renting infrastructure. Google Cloud’s launch post describes deployment to your own Vertex AI endpoints, with compute resources that you provision. Open weights do not mean somebody else pays the electricity bill.
Apache 2.0 Is a Big Part of the Story
Gemma 4’s Apache 2.0 license is commercially permissive. That makes adapting and redistributing a model more approachable, but “permissive” is not synonymous with “no conditions.” The license itself sets out requirements around redistribution, notices and marking modifications, and does not grant blanket trademark rights.
Nor does a model license certify an entire application’s privacy practices. Local inference can keep a prompt on your machine, but an application may still contact external tools, send telemetry or use cloud services. The surrounding software and configuration determine what actually leaves the device.
That distinction matters more than a reassuring label on a download button. If privacy is the reason for going local, check the complete workflow.
A Promising Launch, With Testing Still to Do
The appeal is the combination: competitive open-model results, several sizes and a license that permits substantial freedom to build. None of that requires pretending a launch announcement is our own hands-on review.
The interesting question now is what Gemma 4 can do with the work you actually have. Give it a real task, compare the result with something you trust, and keep the leaderboard in perspective. Third place is encouraging. Correctly handling your particular mess is better.
