EmbeddingGemma 2

Name in the API

google/embeddinggemma-2

Base model

google/embeddinggemma-2

Type

Embeddings

Input → output

text, image, audio, video → vector with 768 dimensions

All inputs share one vector space, so a text query can find images, audio and video.

Shorter vectors with dimensions: 128, 256, 512 or 768 (default).

Context length

8,192 tokens per input, shared by all parts of the input

Availability

Dedicated

Hardware

A 12 GB slice of a GPU per instance

Endpoint

/v1/embeddings

Example

Multimodal search

Notes

• For search, the model authors recommend prefixes for text (not for images, audio or video): queries get task: search result | query: <your query>, documents title: none | text: <your document>.
• Images, audio and video go into input as chat messages, and each message gives one vector. The example shows how.