2026 · 10 · 11 08:16

EmbeddingGemma 2 Explained: Google's Offline Embedding Model for Search

EmbeddingGemma 2 maps text, photos, audio and video into one embedding space at 740M parameters (about 191 MB of RAM for text on a Pixel 11 Pro). What embeddings are, four ways to use it, and how to try Google's Foresight on a Mac.

Open on YouTube
Watch on
The original Chinese episode 谷歌开源 EmbeddingGemma 2：小到普通电脑都能跑，干嘛的？

What this video covers

Chapters

Transcript

A model small enough for a phone

A Google model small enough for a phone can search your files, photos and recordings by meaning, offline. It’s called EmbeddingGemma 2. It’s an embedding model. That means it doesn’t write. It finds. In a few minutes, you’ll know exactly what it does, 4 ways regular people can use it, and how to try it on your own Mac today. Quick refresher. What does an embedding model even do? Your phone can already read the letters in a photo.

What an embedding model actually does

That’s character recognition: pixels in, letters out. Useful, but shallow. An embedding model goes deeper. It turns content into coordinates that stand for meaning. Words go in, numbers come out. Think of every file as a point on a giant map. Similar meaning lands on nearby points. Different meaning lands far apart. Why give things coordinates at all? Because a computer can’t feel that “run” and “jog” mean the same thing. With coordinates, it can measure it. So finding similar things becomes a distance calculation. That’s plain math, and it takes a few milliseconds, even on a phone. Your photo app already does a version of this. It groups pictures by the same face automatically. Embeddings work like that, just for meaning.

It finds, a chat model writes

Here’s the part most people miss. An embedding model only retrieves. It pulls the right material out of a pile and hands it over. A chat model is the one that reads that material and writes a reply. One finds, one writes. Neither can do the other’s job. An embedding model won’t chat with you. It makes sure the chat model finds you.

What’s new in version 2

So what’s new in version 2? The first version only read text. This one maps 4 kinds of content into 1 shared space: text, images, audio, and video frames. A text question can now find a recording. A photo can find a similar photo. No captions or transcripts in between. Google has a demo that makes this click. You type Katze, the German word for cat. Your cat photos appear instantly, before you finish typing. And the results re-rank live as you type more.

How small it really is

Now the size. The whole model is 740 million parameters. Parameters are just the internal settings a model learned. The text-only part is 270 million. Vision and audio are separate add-ons. Don’t need them? Don’t load them. Google squeezed it down and ran it on a Pixel 11 Pro. The text part used 191 MB of memory. Everything together: 567 MB. That’s phone-sized.

Matryoshka: shrink your search index

Here’s a trick worth knowing, called Matryoshka. Normally your content becomes a list of 768 numbers. You can keep only the first 256. Quality stays nearly the same, and your search database shrinks up to 6x. Cutting further, to 128, degrades quality. Test it on your own data first. That warning is straight from Google’s model card.

How good is it? License and downloads

How good is it really? Google calls it the best multimodal embedding model under 1 billion parameters. Their benchmarks even show it beating some specialist models more than twice its size. That’s Google’s own claim, not mine. One number did impress me: code search. It gained nearly 10 points over the previous version. It understands over 100 languages, and it learned from text in over 140. The card is honest: performance isn’t equal across languages. That’s the kind of caveat I like to see. And the license matters. Apache 2.0. Download it, ship it in a product, no fee. Weights sit on Hugging Face and Kaggle today. Local tools like Ollama already support it. The first version passed 20 million downloads, so developers already voted. Google’s own blog spells out the plan: this model is built for private apps that live on your device.

Four ways to use it

Okay. 4 ways normal people can use this. One: search your own stuff by meaning. Files, photos, recordings. Type “that video from the beach last summer” and it pops up. No filename needed. Google’s phone demo finds the moment inside a video, too. You type “dog catching a frisbee” and it jumps to the exact second. Two: feed a private knowledge base. This is the most common real-world use. Ask a question in plain words. The embedding model pulls the relevant paragraphs. Then a chat model writes the answer. That pipeline is called retrieval-augmented generation, or RAG for short. Your company documents never leave the building. Three: group and sort automatically. A new support ticket arrives. Which old tickets does it sit next to? Route it to the right team in milliseconds. The same trick finds you similar articles and songs. Those are my everyday examples, not Google’s demos. Four: search code. Ask for what a function does, in plain words, and it finds the code. That nearly-10-point jump is the biggest upgrade over last year’s model.

Foresight: Google’s Mac demo app

Google also built a demo app so you can touch all this. It’s called Foresight, and it’s a Mac app. Inside, the division of labor we just covered runs live. The embedding model finds material. A chat model called Gemma 4 writes. During a meeting, Foresight listens in the background. You scribble a few words, and it expands them into full notes when the meeting ends. Someone asks a question mid-meeting? It answers on the spot, using the transcript and your imported files. It even pops up a card with the exact source. It works with Meet, with Zoom, and face to face. The knowledge screen is worth a look, too. One user’s screenshot shows over 100 sources imported: documents, PDFs, links, spreadsheets, pictures. A badge says powered by EmbeddingGemma 2. A counter shows millions of tokens, the chunks AI reads, all processed locally. Work that would normally go to a cloud just stays on this one machine.

Nothing gets uploaded

The 3 best words here: nothing gets uploaded. The FAQ spells it out. All inference runs on your computer. Recordings and files never leave the machine. It works on a plane. No subscription, and no cloud costs. Google lists no price on the page.

Three things before you install

3 things to know before you install. First, Apple silicon only. That’s all the FAQ says. Intel Macs and Windows aren’t mentioned anywhere. Second, it’s an experimental demo app, not a polished product. Expect rough edges. Third, the model’s training data ends in January 2025. Ask it about newer stuff and it may not know. Oh, and files in Google Drive must sync locally before Foresight can index them. Offline means local.

My take and your next step

Here’s my take. Embedding models don’t sound exciting, but they’re the piece that makes AI actually know your stuff. Once searching runs on 1 laptop — fast, free, every file type — a personal knowledge base finally makes sense. Foresight is just the first shape you can touch. So here’s your next step. If you have an Apple-silicon Mac, download Foresight. Point it at 1 throwaway folder first. Try it for 1 week before a meeting that matters. Don’t trust it with your real archive on day one. No Apple silicon? Google’s AI Edge Gallery phone app runs the same search demo. If this decoded embedding models for you, subscribe. Every time an open-weight model drops, I explain what it does and who it’s for. Now tell me: would you let a model index your personal files? Yes or no in the comments.

FAQ

What is EmbeddingGemma 2?

Google’s open embedding model released in October 2026. It maps text, images, audio and video frames into one 768-dimensional space so they can be searched by meaning. It has 740M parameters in total and is licensed under Apache 2.0.

What is the difference between an embedding model and a chat model?

An embedding model only retrieves: it turns content into coordinates and finds what is close in meaning. A chat model reads what was found and writes the answer. Systems that combine the two are called retrieval-augmented generation (RAG).

How can I try EmbeddingGemma 2?

On an Apple-silicon Mac, Google’s Foresight app uses it with Gemma 4 for local meeting notes and search, offline. Developers can pull the weights from Hugging Face, Kaggle or Ollama. Start with a throwaway folder before indexing your real files.

Sources

EmbeddingGemma 2embedding modelGoogleoffline searchRAGForesightlocal AI