Gemma 4 Pushes Real AI Onto Android Phones

For years, phone AI has worn a bit of a mask. You ask, it answers, but somewhere in the middle your data slips off to a remote server, gets processed, and comes back to you. That hidden detour has always been the trade-off. Google now wants to cut the cord, and Gemma 4 is the clearest sign yet that on-device AI on Android is getting serious.

Google DeepMind unveiled Gemma 4 last week alongside Arm, with a clear goal in mind: make advanced AI run directly on Arm-based Android phones instead of leaning on the cloud. According to Google, the new model family is up to four times faster than previous versions and can use as much as 60% less battery. The smaller E2B and E4B variants were built with phones in mind, and they can handle text, images, and audio without sending a request over the internet.

Why Arm matters here

The quiet engine behind this jump is Arm’s SME2 instruction set, which is part of newer Armv9 chips. In simple terms, it speeds up the matrix math that modern AI depends on. Arm says its early engineering tests showed an average 5.5x improvement in processing user input, along with response generation that was up to 1.6x faster on the Gemma 4 E2B model. The interesting part? Developers do not need to rewrite their apps to see the gains. Arm’s KleidiAI software layer connects with Google’s existing runtime libraries, so the improvement arrives with very little friction.

That kind of invisible upgrade is exactly what could move on-device AI from a demo into something people actually use every day. Faster. Lighter. Less dependent on a data connection. Those are the three things mobile AI has been chasing from the start.

The clearest example comes from Envision, an accessibility app designed for blind and low-vision users. Until now, scene understanding often depended on cloud access. In a prototype using Gemma 4 locally on Arm CPUs, a user could take a photo and receive a detailed description of the scene instantly, with no network required. For an app like that, offline support is not a bonus feature. It is the feature.

Google is also laying the groundwork for the next step. Gemma 4 is being used as the base for Gemini Nano 4, the upcoming on-device model for Android. That means developers who build with Gemma 4 today should be in line for compatibility with Gemini Nano 4 when it arrives on flagship devices later this year. Gemini Nano already powers local features such as smart replies and audio summaries, and chipmakers like MediaTek have been pushing hard in the same direction. Gemma 4 adds more to the mix, including multimodal support and built-in agentic capabilities.

For developers, access is already open. The E2B and E4B models are available through Google AI Edge Gallery on Android and iOS under an Apache 2.0 license. And that matters because the race for useful, private, low-latency AI is no longer happening only in the cloud. It is moving into the phone in your hand.

Related Post

Read More

Microsoft Edge Finally Fixes a Bizarre Profile Shortcut Gap

Sometimes the smallest software quirks are the ones that make you pause and think: wait, how...