Google puts Gemini on devices with seven 'Gemini Live' agent models for on‑device workflows

Google’s Gemini story is changing. This is no longer just about a chat window where you type a question and get an answer back. The new signal is that Google seems to be wiring Gemini directly into devices, with multiple live voice models and more on-device processing, right before Google I/O on May 19–20. ### What actually showed up? A teardown of Google App v17.18.22 turned up a hidden Gemini Live model selector with seven options that aren’t part of the normal public interface. The interesting part is not just the number. It’s the mix. The leaked options appear to include different Live variants, plus a “Thinking” model and personalization-focused modes, which suggests Google is splitting Gemini Live into specialized behaviors instead of one generic voice assistant. (forbes.com) ### Why does “seven models” matter? Because model choice is really product design in disguise. If Google gives Gemini Live separate brains for fast chat, deeper reasoning, and personal context, then the assistant can quietly swap modes depending on the task. That is how you get from “answer my question” to “help me do something across apps, screens, and time.” In other words, more models usually means more workflow control — not just better small talk. (forbes.com) ### Where does on-device fit in? On-device AI is the part that makes this feel bigger than a feature refresh. Google has already been pushing Gemini Nano as the local model layer for phones and Chrome. Chrome’s Enhanced Protection uses Gemini Nano on-device for scam defense, and Google has been steadily framing Chrome as an AI browser that can understand pages, work across tabs, and move toward agentic browsing. That means the plumbing for local inference is already there. (forbes.com) ### So was the 4GB Chrome model real? The reporting around Chrome quietly downloading a roughly 4GB Gemini Nano model fits the direction of travel, even if the exact rollout details are still a little murky in public documentation. What matters more is the pattern: Google has openly said Chrome is using Gemini Nano locally for security tasks, and it has been expanding Gemini inside the browser as a built-in layer rather than a separate destination. The browser is becoming a host for AI actions, not just a place to open websites. (blog.google) ### Haven’t we seen Gemini Live before? Yes — but mostly as a conversational interface. Over the past year, Google rolled out Gemini Live with camera and screen sharing, first as a premium-style capability and then more broadly across Android and iOS. Google also folded Project Astra’s live visual understanding into Gemini Live. That matters because once a voice assistant can see your screen, see your camera feed, and run some intelligence locally, it stops being just a chatbot and starts looking like an operating layer. (blog.google) ### Why is Google doing this now? Because competitors already moved. OpenAI and Anthropic have been pushing computer-use and agent-style systems that can take actions, not just generate text. PCWorld’s read on the pre-I/O clues is basically that Google is trying to catch up fast by making Gemini proactive, persistent, and embedded across Chrome, Android, and the app stack. Seven hidden Live models would fit that exact strategy. (blog.google) ### What should we watch at I/O? Watch for Google to connect three things in one story: Gemini Live, Chrome, and Android. If Google formally announces multiple Live modes, more local Gemini Nano usage, and deeper app or browser control, then the company is setting a new baseline for what a default assistant is supposed to do. The catch is that this also gives Google more power over how tasks get routed, summarized, and completed inside its own ecosystem. (pcworld.com) ### Bottom line The real news is not “Gemini got more models.” It’s that Google appears to be turning Gemini into device software — always there, increasingly local, and much closer to acting on your behalf than waiting for a prompt. (forbes.com)