AI Edge Gallery on Mac Review: Local Gemma 4, Zero Setup

Google shipped a Mac version of AI Edge Gallery on June 4, and it changes who gets to run a local language model on Apple hardware. As AppleInsider reported, this is the first time Google has offered its own tool for running Gemma models locally on the Mac: download the app from Google’s developer site, pick a model, and everything runs offline. You never see a terminal, and there is nothing to configure.

The short version of our take: if you have been curious about local LLMs but never wanted to learn Ollama’s command-line workflow, this is now the cheapest possible entry point on a Mac. If you already run local models through a general-purpose manager, the app itself gives you little reason to switch, though one of its companion releases might.

What AI Edge Gallery actually is

AI Edge Gallery is Google’s consumer-facing showcase for on-device AI, and until this month it was a phone-only product. A Google blog post from June 9 still pointed users to “Google AI Edge Gallery on iOS or Android” as the way to try Gemma 4, which tells you how recent the desktop gap was.

The Mac release closes that gap. Per AppleInsider, the app became available as a direct download from Google’s developer site on June 4. Power users could already run Gemma on a Mac through third-party tooling, so the news here is less about capability than packaging: an official, GUI-first path from Google itself.

The same report lays out why local execution matters in the first place. The model works without an internet connection and your prompts never leave the machine. Skipping the round trip to a cloud server also tends to make responses faster. None of that is unique to Google’s app, but bundling it behind a one-click installer is what moves local AI from a hobbyist project to something you can recommend to a colleague.

Five models in the box, one memory line

The Mac app ships with five Gemma variants. AppleInsider’s list breaks down like this:

Model	Role
Gemma-4-12B-it	New flagship, built for agentic multimodal work on laptops
Gemma-4-E4B-it	Edge tier, the middle option
Gemma-4-E2B-it	Edge tier, lightest Gemma 4
Gemma-3n-E4B-it	Previous-generation edge model
Gemma-3n-E2B-it	Previous-generation edge model, smallest footprint

The dividing line is memory. Google puts the floor for Gemma 4 12B at 16GB of VRAM or unified memory, and AppleInsider notes that every modern Apple laptop clears that bar except the MacBook Neo. An 8GB machine can still use the app; it just lives in the E2B and E4B tiers instead of the flagship.

That tiering logic will feel familiar if you read our Qwen3.5 small-model review, where the conclusion was that your RAM, not your enthusiasm, decides which local model you actually run. Google has effectively turned that decision into a dropdown menu.

The model is the real headline

The app is the delivery vehicle. The cargo is Gemma 4 12B, which launched the same day, and the architecture story is worth a closer look.

Chrome Unboxed’s breakdown highlights two numbers. The 12-billion-parameter model delivers multi-step reasoning that lands close to Google’s own 26B Mixture of Experts model, while using less than half the memory, and it is tuned specifically for consumer laptops with 16GB of RAM.

The more unusual claim is how it handles images and audio. Conventional multimodal models route those inputs through separate encoders before the main network sees them, and those extra stages cost latency and memory. According to Chrome Unboxed, Gemma 4 12B uses an encoder-free design in which visual and audio signals feed straight into the core backbone. That makes it Google’s first mid-sized open model with native audio input, able to transcribe, clean up, and translate speech entirely offline.

On licensing and scale, Google’s own blog says the Gemma 4 family ships under Apache 2.0, has passed 150 million downloads, offers up to a 256K context window on the larger models, and now has quantization-aware training checkpoints available. We argued in our Gemma 4 launch analysis that Google treats its open models as a land grab for local agent workloads. Putting AI Edge Gallery on the Mac extends that grab from phones to desks.

Is this an Ollama replacement?

No, and it does not try to be.

AI Edge Gallery is a curated storefront. Google decides the catalog, the catalog is Gemma-only, and in exchange you get zero setup and models pre-tuned for your hardware. Ollama and LM Studio are warehouses: any open model you want, any quantization you choose, with all the selection and tuning work left to you.

So the question to ask is not which tool is better but which problem you have. Want a free, offline assistant on a Mac with no learning curve? The Gallery wins on convenience. Want to run Qwen, Llama, or anything outside Google’s family, including the small models we compared in the Qwen3.5 review? The Gallery cannot help, and the general-purpose managers remain the only road. The two categories coexist rather than compete.

Developers get a third angle: the Gallery works as a free compatibility lab. Google’s builders showcase describes how HubX built BetterSpeak, an offline English-tutoring app, on a 4-bit quantized Gemma 4 E2B, processing all voice and text on the device. If that is the kind of product you are planning, the Gallery lets you verify what each Gemma tier can do on real hardware before you commit to an integration.

The supporting cast: Eloquent and DiffusionGemma

Two more releases this week fill out the picture of where Google is pushing on-device AI.

AI Edge Eloquent arrived alongside the Mac app. AppleInsider describes it as an on-device dictation and text-editing tool that you summon with a keyboard shortcut and use inside any Mac application, with no connection required. It was previously iPhone-only. For anyone who drafts by voice or transcribes meetings, this utility may end up getting more daily use than the Gallery itself.

DiffusionGemma followed on June 10. Google’s announcement details an experimental 26B MoE model that activates only 3.8B parameters at inference and generates whole blocks of text at once through diffusion instead of token-by-token decoding. Google quotes up to 4x faster generation on dedicated GPUs, north of 1,000 tokens per second on an H100, with the quantized model fitting in 18GB of VRAM. The company is upfront that output quality trails standard Gemma 4. This one is a research toy for speed-critical interactive work, not a daily driver.

Set these next to the agent-platform strategy from Google I/O 2026 and the division of labor is plain: Gemini owns the cloud, Gemma owns the device, and AI Edge Gallery is the consumer doorway into the second half of that split.

Verdict: who should install it

If you own a Mac and have never touched a local LLM, install it. The app costs nothing and asks for no setup knowledge. Start with the 12B model if you have 16GB or more of unified memory, and drop to E2B on older 8GB machines.

If you already run Ollama or LM Studio, skip the Gallery and take Eloquent. The app adds nothing to a workflow that handles arbitrary models, but a free system-wide offline dictation tool is worth grabbing on its own.

If you are a developer eyeing on-device features, treat it as a test bench. Confirm how each Gemma tier behaves on the machines your users own, then follow the BetterSpeak quantization playbook when you build.

One gap remains worth flagging: the desktop release covers only the Mac, and Windows users are still waiting. Given how Google staged the rollout from phones to Apple’s desktop, a Windows build feels like a matter of sequencing rather than intent. Until it lands, the path to local models on Windows still runs through the general-purpose tools.