Unlock local, agentic workflows with Gemma 4 12B and Google AI Edge, directly on your laptop. Experience 100% on-device AI: • Generate code in AI Edge Gallery (new to Mac) • Dictate and edit text via AI Edge Eloquent (new to Mac) • Serve Gemma 4 12B locally with... LiteRT-LM Dive in:show more

👀 Gemma 4 12B running fully on-device is the quiet shift everyone's been sleeping on Next stop: local agents that never touch the cloud 💻

Going beyond the cloud straight to the Gemmaverse!

Let's gooo

100% on-device AI keeps getting more practical. A 12B model running locally would've sounded wild not that long ago.

Local Gemma 4 12B is a game-changer for dev workflows. No API calls, no latency, no data leaving your machine. Perfect for prototyping AI agents that need to run anywhere. The code generation alone makes this worth testing.

只是 AI 性能比较堪忧

please, add Gemma 4 12B to Edge Gallery for android

'Local + agentic' is the right push — but on-device the wall is memory bandwidth, not params. Every token moves the active weights through memory once. A phone at ~70 GB/s caps 12B-INT4 near ~12 tok/s — below agent speed. The small encoder-free variants are the real edge play.

Local AI gets useful for normal people when it handles repeatable private tasks: drafts, notes, simple code checks, document cleanup. The model news is cool, but the daily workflow question is: what can safely run on my laptop without sending everything away?

100% on-device means your data never leaves your laptop. OpenAI can't train on what it can't see. The real disruption isn't cheaper tokens, it's cutting the cloud out entirely.

When is the windows version gonna be available

Did the guy record in an airplane?

on-device codegen on a laptop changes behavior, not just benchmarks. no round-trip, no token meter running, no "is this leaving my machine" — you just iterate. local + agentic is where privacy-sensitive devs quietly move first. trying AI Edge Gallery now.

this isn't agentic workflow : )) it's just an offline chatbot

@GoogleAI I use AI Edge Eloquent regularly. While I understand the dictation accuracy will go up over time, when it gets it wrong and you give an option to edit the transcribed text, you should not run the refinement again, that makes it worse and the text becomes useless.

Can't wait for the Ollama release! Running Gemma 4 12B locally on Linux with a 16GB RAM laptop will be the perfect setup for a local agentic assistant!

That is amazing to try…

Impressive update! I'm curious about the actual hardware requirements for running Gemma 4 12B smoothly on a Mac laptop. Has anyone tested the RAM usage yet?

This is fantastic! Love seeing Eloquent finally on Mac with hotkey support! On iOS though, the biggest missing piece is still a proper custom keyboard extension (or at least a floating mic button like the early App Store listing hinted at). Right now we have to open the full app → dictate → copy → paste, which kills the flow for everyday use. A system-wide keyboard would make this genuinely indispensable on iPhone/iPad. Any plans to bring that to iOS? Huge thanks to the team for the fully on-device Gemma-powered magic either way! 🙌

💖

@googledevs This has been incredible, but I can't get it to work in Obsidian. Instead of typing what I say, it activates the voice editing command mode.

@grok will this work with M3 Maxbook Air wirh 24 GB RAM

Why do all the fire AI features lowkey drop on Mac first? 🙄 Like, when is Windows OS finally getting some love? Google is straight up gatekeeping at this point, it’s not it! 💀 #WindowsUser #GoogleGemini

What are the application scenarios for this?

Do you all even QA your products? I’ve installed this on two different M4 MacBook Pros, and it consistently hangs when “polishing text” using the Gemma 4 12B model. When I try to switch models, it hangs there too. Back to Wispr Flow . . .

More info here

😭😭 Windows 😭😭

dont give a fuck about mac, give us windows

google ai edge能操控app吗？

What about a Windows version?

Local agents feel like going from paper trading to live account access. Cloud agents are one thing. But once it runs on your laptop, the question changes: what can it touch, when should it stop, and who takes the hit if it goes wrong?

Where we can find the trascribed file ? As i recoreded one session the. The file was no where to found plz help

My ai eloquent doesnt paste into the texting field when i dictate it simply copies to clipboard, i have tried to change it in the setting but it has not changed. What can i do?

On-device agentic is the underrated shift - latency and privacy stop being trade-offs once the loop runs locally. The constraint that matters isn't model size, it's how tight you make the plan-act-eval cycle. Built something similar -