Google releases EmbeddingGemma 2, a compact multimodal search model for developers

Google has introduced EmbeddingGemma 2, a new open-source embedding model built to handle search and retrieval-augmented generation (RAG) across multiple content types. The sub-1B model, based on Gemma 4 and released under the Apache 2.0 license, aims to bring multimodal retrieval capabilities to applications without requiring massive computational infrastructure.

What EmbeddingGemma 2 does#

The model bridges a persistent gap in building modern search applications. Rather than cobbling together separate tools for text, code, images, video, and audio, EmbeddingGemma 2 processes all these formats through specialized encoders that project into a single 768-dimensional vector space. This unified approach simplifies architecture and enables direct semantic matching across modalities—a text query can be scored against images, audio clips, or code snippets using the same index.

Performance gains over its predecessor are substantial: the new version scores 14% higher on code retrieval benchmarks and adds visual document understanding (PDFs, slides, charts) alongside native video and audio support.

Modular and memory-efficient design#

A distinctive feature is granular loading control. Developers can choose which encoders to activate:

• Text and code only (270M parameters) — the lightweight base
• Text + vision (440M parameters) — adds image and video understanding
• Text + audio (570M parameters) — adds speech and sound processing
• Full multimodal (740M parameters) — all four modalities enabled

Only the selected encoders load into memory, lowering the barrier to deployment on resource-constrained devices or commodity servers. All four configurations share the same vector space, so embeddings computed with one setup remain compatible with another.

The model also supports Matryoshka Representation Learning (MRL), which allows dynamic truncation of vectors from 768 dimensions down to 128. A truncated 256-dimensional embedding retains roughly 95% of retrieval quality for images, video, and speech while requiring only a third of the storage.

How to use it#

Developers can integrate EmbeddingGemma 2 via the sentence-transformers library (v6.1.0+) with a simple pip install. The model supports task-specific prompts—for instance, encoding queries with the "SearchQuery" prompt and documents with the "Document" prompt optimizes retrieval accuracy.