Google's On-Device AI Push Is Real — And Your Phone Is the Proof
2026-04-06 • AI & Mobile • Butler
Google is making its most concrete push yet to put working agentic AI on phones — not chatbots, not voice assistants, but actual multi-step autonomous workflows running entirely on your device. Here is what Gemma 4 actually delivers on mobile.
The story most Gemma 4 coverage has missed is the phone story.
Most articles have focused on benchmark numbers, the Apache 2.0 license, or the open-versus-closed debate. Those things matter. But Google's more interesting move this week is quieter: it shipped actual on-device agentic AI to phones — not a demo, not a concept, but working code you can run offline on an Android device today.
That is worth understanding separately from the broader Gemma 4 launch narrative.
What Google actually shipped to phones
Gemma 4 comes in four sizes. The two that matter for phones are the E2B (Effective 2B) and E4B (Effective 4B) — smaller models optimized specifically for edge and mobile hardware. They are not the most powerful models in the family. They are the most efficient.
Google built these in close collaboration with Google Pixel, Qualcomm, and MediaTek. The design constraint was not maximum benchmark score — it was how well the models run on actual phones with actual battery constraints.
The result: E2B and E4B run completely offline with near-zero latency on phones, Raspberry Pi, and NVIDIA Jetson Orin Nano. They support text, images, and audio inputs. They handle over 140 languages. And they are available now through two paths:
AICore Developer Preview — Google's system-level AI framework on Android — now includes Gemma 4 E4B and E2B as downloadable models. If you have a supported Android device, you can run these models directly on the hardware without an internet connection.
AI Edge Gallery — a mobile app available on iOS and Android — launched with a feature called Agent Skills, which runs multi-step autonomous workflows on-device. The example Google gives: an agent that can query Wikipedia, summarize content, and generate visualizations entirely through local model calls, with no cloud involved.
That is the part that usually gets lost in "on-device AI" coverage. Most phone AI still phones home for the hard stuff. This is designed not to.
What Agent Skills actually do
Google's AI Edge Gallery Agent Skills are worth describing specifically, because the "on-device AI" label gets applied to everything from voice transcription to full agentic loops.
Agent Skills, as Google describes them, can:
Query Wikipedia and knowledge sources and return synthesized answers
Generate summaries, flashcards, and data visualizations from input
Integrate with other on-device models — for example, pairing photo understanding with text-to-speech or music synthesis
Build end-to-end experiences through multi-step conversation rather than single-prompt exchange
The last point is the meaningful one. A single-prompt model that runs locally is useful. An agent that can plan across multiple steps, call tools, and complete a task without a round-trip to a server — that is the harder problem. Google is claiming E2B/E4B can do the latter.
That claim is worth treating as a starting point, not a verified benchmark. But it is a more concrete starting point than most "on-device AI" announcements have offered.
The forward path: what is still coming
Gemma 4 E2B and E4B are explicitly described by Google as the base for the next generation of Gemini Nano 4 — the version of Google's on-device model that will ship to mainstream Android users, not just developers in a preview program.
Google's stated claims for that next version: up to 4x faster inference and up to 60% less battery use compared to the current Gemini Nano generation. Those are Google's forward-looking product claims, not independent benchmarks. The actual production experience will depend on device, workload, and implementation.
What matters is the trajectory. Google is using Gemma 4 as the development and testing ground for what eventually becomes the standard on-device experience for Android. That means the AICore Developer Preview today is a preview of a mainstream product in the not-distant future.
Why the mobile angle is the real Gemma 4 story
The open-model vs. closed-model debate is real, and Gemma 4 is a meaningful entry in that space. But for most people who use Android phones, the more immediate question is: what can my phone actually do with AI that does not require sending my data somewhere else?
The honest answer as of this week: limited but real.
You can run Gemma 4 E2B/E4B on a supported device right now through AICore Developer Preview. You can experiment with Agent Skills through AI Edge Gallery. You can prototype local AI features in Android Studio using the ML Kit GenAI Prompt API.
What you cannot do yet: run the full production version of Gemini Nano 4. That is still coming. And when it arrives, Google's 4x speed / 60% battery improvement claim will be worth testing against real-world conditions, not just accepting as fact.
The on-device AI story is real. It is also still in progress.
Butler take
Google has made a more credible on-device agentic AI push than anything it has shipped before. E2B and E4B are genuinely designed for phones, the AICore Developer Preview is a real shipping path, and Agent Skills in AI Edge Gallery are a concrete example of what local agentic workflows look like.
The gap between "Google announced this" and "your phone runs this today" is still real, and the forward path to Gemini Nano 4 is a stated claim, not a delivered product. But the direction is clear, and the building blocks are more serious than the hobbyist demos that "on-device AI" usually conjures.
If you have a supported Android device and want to see what local agentic AI actually looks like today, the AICore Developer Preview and AI Edge Gallery are the place to start.
---
Related coverage
AI Disclosure
This article was produced with AI assistance for research synthesis and drafting, then reviewed and edited for clarity and accuracy. Claims about future products (Gemini Nano 4 speed and battery improvements) are presented as Google's stated forward-looking claims, not independently verified benchmarks.