# Introducing EmbeddingGemma 2: An open model for natively multimodal embeddings

## Executive summary

Google DeepMind introduced EmbeddingGemma 2, a lightweight, open model designed for natively multimodal embeddings. This model maps text, images, video, and audio into a single unified embedding space, enabling comprehensive search and retrieval (including video moment finding) entirely offline on edge devices. With a maximum of 740 million parameters and an 8,000 token context window, it supports building private Retrieval-Augmented Generation (RAG) pipelines, even when paired with generative models like Gemma 4, ensuring sensitive data never leaves the hardware.

## Key takeaways

- Multimodal Embedding Capability: EmbeddingGemma 2 unifies cross-modal retrieval by mapping text, images, video, and audio into a single shared high-dimensional embedding space, allowing a single model to handle search, retrieval, and classification across diverse data types.
- On-Device Efficiency and Architecture: The model is engineered for edge hardware, featuring a modular form factor with a maximum of 740 million parameters. It outputs 768-dimension vectors, which can be truncated down to 128 dimensions using Matryoshka representation learning. It supports an 8,000 token context window for embedding long documents or code bases.
- Offline and Private Pipelines: The model enables instant media search (e.g., finding specific moments in a video) and powers RAG pipelines (e.g., in the AI Edge Foresight app). Because all embedding and processing occurs on the device, sensitive user data remains private and no external API calls are required.
- Customization and Deployment: While offering strong out-of-the-box quality, the model can be fine-tuned for domain-specific vocabularies (e.g., legal contracts, medical imaging) to improve retrieval precision without increasing model size.

## Technical details

- Model Specifications: EmbeddingGemma 2 has a maximum of 740 million parameters and outputs 768 dimension vectors, with the ability to truncate to 128 dimensions via Matryoshka representation learning. It features an 8,000 token context window.
- Multimodal Mapping: The model maps text, images, video, and audio into a single unified embedding space, allowing for unified search and retrieval.
- Application Use Cases: Use cases include Instant Media Search (retrieving photos/assets via natural language queries) and Video Moment Finder (finding specific segments within videos).
- System Integration: It can be paired with lightweight generative models like Gemma 4 to build RAG pipelines and agentic workflows, ensuring data privacy by keeping processing local to the hardware.

## Practical implications

- Enables building complex, privacy-preserving AI pipelines (RAG) that operate entirely offline on edge devices.
- Reduces the need for separate, specialized models for different data modalities (text, image, video, audio).
- Allows for efficient embedding of large datasets (up to 8,000 tokens) directly on the device for local search and retrieval.
- Provides a foundation for domain-specific AI applications by allowing fine-tuning on specialized vocabularies.

## Topics

Multimodal AI, Edge Computing, Embeddings, RAG, Generative AI, Learn about the model, Developer Guide, Google AI Edge, HuggingFace (Model Weights), Gemma Cookbook

Source: https://www.youtube.com/watch?v=anPsS6huQk0