AI BriefingPickembeddinggemma-2: 740M parameter model integrating text, image, and audio into a single vector
google/embeddinggemma-2
About the project
A multimodal embedding model that maps text, image, video, and audio into a single 768-dimensional vector space. With a total of 740M parameters, it features a structure combining a 270M text backbone with modular vision and audio encoders.
It allows selective loading of only the required modalities, enabling low-latency inference on consumer hardware such as mobile devices and laptops. It supports Matryoshka Representation Learning, allowing flexible adjustment of vector dimensions from 128d to 768d, reducing storage costs by up to 6x.
It understands over 100 languages and code, with code-related performance improved by approximately 14% compared to the previous generation. With an 8K token context window, it supports processing of audio and video lasting several minutes.
Task steering using lightweight text prefixes is possible for various tasks such as search, RAG, classification, and clustering. Distributed under the Apache 2.0 license, it is suitable for building on-device applications.
google/embeddinggemma-2
The original page has no description.
feature-extraction
This introduction was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report errors, attribution issues, or removal requests via Contact.