Google just released Gemma 4, and it's climbing leaderboards fast. The 31B dense variant is now the #3 open model worldwide on the Arena AI text leaderboard. The 26B mixture-of-experts model ranks #6. And unlike most frontier-capable models, it's completely free under the Apache 2.0 license.
Let that sink in. A genuinely capable AI model - one that can compete with commercial offerings - is now available for anyone to download, modify, and deploy without restrictions or API costs.
Gemma 4 comes in four sizes: E2B, E4B, 26B MoE, and 31B Dense. All models natively process video and images with variable resolutions. The E2B and E4B variants have native audio input capabilities. This is multimodal AI that runs locally, not in the cloud.
Google explicitly positioned Gemma 4 for on-device agentic intelligence on Android. The smaller variants are designed to run on phones. Think about that: frontier-capable AI models running entirely on your smartphone, with no network connection required.
One developer commented: "Already running the 26B MoE on my RTX 4090. It's remarkably good for its size. Purpose-built for reasoning and agentic workflows."
This is the future that was science fiction two years ago. Locally-run AI that's actually capable, not just a toy.
The Apache 2.0 license is key. It's more permissive than many other "open" AI models, which often have commercial use restrictions or require attribution in ways that make deployment complicated. Apache 2.0 means you can use Gemma 4 in commercial products, modify it, fine-tune it, and deploy it however you want.
Why would Google give this away? Two reasons:
First, ecosystem growth. By open-sourcing capable models, Google encourages developers to build applications on their architecture. Those applications create demand for Google Cloud services, Android devices, and the broader Google ecosystem. It's the Android playbook applied to AI: give away the platform, monetize the services.
