Google Releases EmbeddingGemma 2: A 740M Multimodal Embedding Model for Mobile Devices
Google has launched EmbeddingGemma 2 under an open-source Apache 2.0 license, introducing a modular 740-million-parameter multimodal model that maps text, code, image, video, and audio into a shared vector space. When quantized for edge devices, the model requires as little as 191 MB of RAM for text-only tasks and around 567 MB for full multimodal capability. This release enables privacy-centric, fully offline cross-modal search and Retrieval-Augmented Generation (RAG) directly on consumer smartphones without relying on cloud infrastructure. Its low memory footprint and seamless interoperability with Gemma 4 significantly reduce deployment costs and pipeline latency for mobile developers. The model incorporates Matryoshka Representation Learning (MRL), allowing developers to dynamically truncate output embeddings from 768 to 128 dimensions to shrink vector database footprints up to 6x. It also offers an 8K token context window capable of processing up to 5.5 minutes of audio or 58 video frames in a single prompt.
## BACKGROUND
Embedding models convert unstructured data like text, audio, and images into numerical vectors so computers can quickly calculate similarity for search and retrieval tasks. Benchmarks such as MTEB (Massive Text Embedding Benchmark) measure embedding quality across diverse retrieval tasks, while Matryoshka Representation Learning (MRL) allows vectors to be sliced to smaller sizes without retraining.