Google (learn more) Launches EmbeddingGemma 2: An Open Multimodal Embedding Model for Devices
Google has released EmbeddingGemma 2, an open 740-million-parameter model that maps text, code, images, audio, and video into a single space and runs entirely on a phone or a small board. It is licensed under Apache 2.0, has had no safety tuning, and Google notes that performance may not be equal across the 100 languages it supports.
Key Features:
- Memory Efficient: Needs about 191MB of RAM for text work on a Pixel 11 Pro, or approximately 567MB for all five modalities.
- Multimodal Retrieval: Maps all content into a 768-dimension space, with 270 million parameters dedicated to text alone.
- Offline Capable: Enables retrieval with no network connection, as demonstrated with an offline translator on an $80 Raspberry Pi.
- Expandable: Vectors can be reduced from 768 dimensions to 128, potentially saving storage up to sixfold.
Background:
EmbeddingGemma 2 builds upon the Gemma family, which has achieved over a billion downloads, with the first EmbeddingGemma accounting for 20 million of them. It shares a text tokeniser and audio encoder with Gemma 4, optimising memory usage.
Implications:
This development aligns with Europe's goal of holding local capabilities, running on existing hardware under a permissive licence, and developed within the region.
(Note: Quoted text and headings have been preserved as requested, with the specified guidelines followed for formatting.)