Sound, photos and text, in any language, in one space
EmbeddingGemma 2 turns a sound, a photo or a sentence in any language into 768 numbers. Things that mean the same land close together, so a dog's bark ends up next to “perro ladrando”. Pick an example or try your own.
Try an example · or put in your own below
Why not just cosine similarity?
Smaller vectors
The model is trained Matryoshka-style, so its first 512, 256 or 128 numbers still work on their own.
Words come from the model's own vocabulary (91,194 words in every language). Green tags mark other languages, with the nearest English word after “=”. Tap any word to explore it.
For the curious Where A and B land, any word's neighbours, and the 768 numbers themselves.
Where they land
Find a word, or tap one on the map or in A or B.
Inside a vector
Put something in A or B, or find a word, to look inside its vector.
Indigo positive, rose negative; stronger colour = higher rank by size in this vector.
Strongest dimensions
Every value
Looking it up…
Rogue dimension: large for nearly every word, so it says little about meaning. The words furthest above and below its average are more telling.
How a recording, photo or text becomes numbers, step by step: How it works.