EmbeddingGemma 2, explained
A 5-minute visual guide

One space for text, photos, video & sound

What an embedding is, what makes it good, and how Google’s EmbeddingGemma 2 takes CLIP’s 2021 idea of a shared vector space and extends it to every sense. All numbers on this page come from a real run on a laptop CPU.

768 dimensionstext · codeimages · videoaudioApache 2.0
01
What is an embedding?

Meaning, written as numbers

An embedding is a list of numbers that captures what something means. A model reads the input and places it at a point in a space with 768 directions. Things that mean similar things end up close together.

Input
EmbeddingGemma 2ENCODER
Output · 768 numbers
Like a colour code: orange→R 240 · G 90 · B 20 = 3 numbers that sit near other oranges. An embedding does the same for meaning, with 768 numbers instead of 3.

These are the real first values the model produced. Each square is one number: orange = positive, blue = negative, darker = larger. No single number means “dog”. The meaning lives in the whole pattern.

02
What makes an embedding good?

Similar meaning → closer vectors

A good model puts a sentence and its paraphrase close together, even when they share almost no words, and puts unrelated sentences further away. “Closeness” is measured by cosine similarity, the angle between two vectors: the smaller the angle, the closer to 1.

Angle between vectors (to scale)

A and B share no key words (dog / puppy, beach / sand by the sea), yet score . A and C score . Raw scores aren’t percentages: with this model even unrelated text sits fairly high. What matters is the ranking, and the paraphrase clearly wins.

03
What is EmbeddingGemma 2?

One open model, one space for every kind of input

Google DeepMind’s EmbeddingGemma 2 (6 Oct 2026) puts text, images, video and audio into the same 768-number space, in a model small enough to run on a laptop or phone.

Wasn’t this around before? Partly. The idea grew step by step
04
CLIP vs EmbeddingGemma 2

Same paradigm, wider reach

CLIP (OpenAI, Feb 2021) showed that two encoders trained on 400 million image–caption pairs can share one space, so “a photo of a dog” lands near dog photos, with zero-shot classification as a result. EmbeddingGemma 2 keeps that joint-space idea and widens it.

CLIParXiv:2103.00020 · 2 senses
Imagepixels Texta caption Image encoderViT or ResNet Text encoderTransformer Shared spaceimage ≈ captionfixed size vectorvector Training: match each image to its own caption two towers meetonly at the very end
EmbeddingGemma 2Google DeepMind · Oct 2026 · 4 senses
Text · Codetokens Images · Videophoto · frames Audio16 kHz wave Vision encoder167M Audio encoder305M Gemmabackbone271Mone sharedtransformer Joint spacemean-pool → 768-dnormalized tokenstokens Truncatable vector Mixed input works too: one vector for text + photo + audio together (placeholders <|image|> <|audio|> <|video|> inside the text)
CLIP → GEMMA 22 senses → 4image + text becomes text/code, images, video and audio in one space.
CLIP → GEMMA 2Two towers → one backbonevision and audio encoders turn media into tokens that one shared Gemma transformer reads.
CLIP → GEMMA 2Fixed size → truncatablekeep the first 512, 256 or 128 numbers (Matryoshka) to store less.
In practice: whatever goes in, the same 768-number format comes out
And they’re comparable: “” vs every photo

The sentence was never paired with any of these photos. It still picks the right one, ahead of the next. That’s the joint space at work.

05
Takeaway

Same core idea, more senses, and it runs where your data lives

CLIP showed that photos and text can share one vector space. EmbeddingGemma 2 keeps that idea and adds audio and video: every input becomes the same kind of 768-number vector. The model is small and open, so the whole pipeline, from embedding your files to searching them, runs on your own laptop, phone or server. Your files never have to leave it.

🔒 On your device — nothing is uploaded
1 · Your data
📷 photos
🔊 sounds · 🎙 voice notes
💬 chat messages
2 · Embed once
EmbeddingGemma 2
each file → 768 numbers
3 · Vector index
768 numbers per file
optional: keep only the first 256 to save space
4 · Search
“my dog at the beach”
→ 768 numbers → nearest items
5 · Results
📷 beach photo
💬 Bruno chat
🔊 the bark
Try it: the sample lab

A small “phone” with 20–30 everyday photos, sounds, voice notes and chat messages, all embedded with EmbeddingGemma 2. Type any sentence and get back the closest items of every kind, ranked, with no keywords, tags or captions.

Open the lab →