Google has unveiled EmbeddingGemma 2, an open-source model with 740 million parameters that consolidates text, code, images, audio and video into a unified vector space while operating entirely on mobile devices and small computing boards. The model requires approximately 191MB of memory for text-only operations on a Pixel 11 Pro, or around 567MB when handling all five modalities. Distributed under the Apache 2.0 licence, the model comes without safety tuning and performs unevenly across the 100-plus languages it supports.
The architecture maps all content into a single 768-dimensional space. The model comprises 740 million parameters overall: 270 million dedicated to text processing, a 170-million-parameter vision encoder and a 300-million-parameter audio encoder that load only when needed. By keeping processing local to the device, the entire retrieval pipeline operates without network connectivity.
Embedding models function by converting content into numerical representations, enabling software to locate information based on semantic meaning rather than keyword matching. This computational step traditionally occurs on remote servers. EmbeddingGemma 2 builds on the architecture introduced with Gemma 4, which arrived in April as four open-weight models under the same licence, designed to run on phones, Raspberry Pi boards and consumer graphics cards. The new embedding model shares the text tokeniser and audio encoder from that family, allowing both to operate together with reduced combined memory requirements.
Google advanced a similar rationale in August when it deployed an offline translation system on an $80 Raspberry Pi for managing sensitive conversations without transmitting audio beyond the device. Maintaining data locally circumvents the transfer step, which represents the primary entry point for European data protection considerations.
The Gemma family has surpassed one billion downloads, with the original EmbeddingGemma accounting for 20 million of those. The successor model accommodates 8,192 tokens across all modalities—quadruple the previous version's capacity—translating to 29 images, 58 video frames or five and a half minutes of audio. Vector dimensions can be reduced from 768 to 128, which Google indicates can lower storage requirements by up to sixfold.
Trade-offs and limitations
The model operates without safety tuning or output moderation; instead, mitigation measures were applied during training data preparation. Support extends to more than 100 languages, though Google acknowledges that performance may vary significantly across different language implementations. The EU recognises 24 official languages.
The capability represents precisely the kind of infrastructure Europe has repeatedly expressed interest in maintaining locally—running on hardware already owned by European companies, under a licence available to anyone—yet it originated from Mountain View.
Source: The Next Web
