Introducing EmbeddingGemma 2: An open model for natively multimodal embeddings
Read the Portuguese translation →
Author: Google for Developers · Watch on YouTube
Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.
Original language: English
Original video: Introducing EmbeddingGemma 2: An open model for natively multimodal embeddings · Channel: Google for Developers · 02:39 · Watch on YouTube
Topics: Data, databases and embeddings, Local and open-source models, Privacy and data protection, RAG, search and information retrieval
Summary
Section titled “Summary”EmbeddingGemma 2 is presented as a lightweight, open model that maps text, images, video, and audio into one shared embedding space. It supports search, retrieval, classification, and clustering, with a modular design of up to 740 million parameters and options to reduce embedding dimensions. On-device examples include searching photos and video moments, as well as powering private RAG pipelines and agentic workflows. The model can also be fine-tuned for domain-specific vocabularies and assets, and its weights and notebooks are available to explore.
Key points
Section titled “Key points”- EmbeddingGemma 2 maps text, images, video, and audio into a unified embedding space.
- The shared space enables search, retrieval, classification, and clustering across different data types.
- The model has a modular form factor with a maximum of 740 million parameters.
- Its 768-dimension vectors can be truncated to as few as 128 dimensions using Matryoshka representation learning.
- An 8,000-token context window supports embedding long documents, code bases, and extended voice recordings.
- Google AI Edge Gallery demonstrates natural-language search across photos and video moments directly on a phone.
- On-device processing can work offline without intermediate transcription or external API calls.
- Paired with lightweight generative models, it can support RAG pipelines and agentic workflows while keeping sensitive data on the hardware.
- Fine-tuning can target domain-specific vocabularies and assets without increasing model size.
Chapters
Section titled “Chapters”| Start | Chapter | Summary |
|---|---|---|
| 00:00 | One shared space for multimodal embeddings | EmbeddingGemma 2 maps text, images, video, and audio into a unified embedding space for cross-modal applications. |
| 00:33 | Model capabilities and efficiency | The video describes the model’s benchmark performance, modular size, adjustable embedding dimensions, and long context window. |
| 01:05 | On-device media search | Examples show natural-language searches retrieving photos and specific moments in videos directly on a phone. |
| 01:38 | Private local workflows | The video demonstrates how on-device embeddings can support RAG and agentic workflows while keeping sensitive data local. |
| 02:08 | Fine-tuning and getting started | EmbeddingGemma 2 can be fine-tuned for specialized assets, and viewers are directed to download its weights and explore notebooks. |
Quotes
Section titled “Quotes”into a single unified embedding space. — 00:00
you can build search, retrieval, classification, and clustering for different types of data. — 00:33
So my entire search can happen completely offline. — 01:38
sensitive user data never leaves the hardware. — 02:08
This helps improve retrieval precision without increasing model size. — 02:39
Concepts and entities
Section titled “Concepts and entities”Embedding Gemma 2, cross-modal retrieval, Matryoshka Representation Learning, Google AI Edge Gallery, Gemma 4, AI Edge Foresight, Hugging Face, Gemma Cookbook
Full transcript (original language)
Section titled “Full transcript (original language)”00:00 Today, we’re super excited to introduce EmbeddingGemma 2, a lightweight, open model for natively multimodal embeddings. - Traditionally, cross-modal retrieval required separate models, one for text, another for vision, another for audio. EmbeddingGemma 2 changes that by mapping text, images, video, and audio, or any combination of them into a single unified embedding space. This means a picture of a cat, the word cat, and the sound of a cat are all mapped close to each other
00:33 in a shared high-dimensional embedding space. So with this one model, you can build search, retrieval, classification, and clustering for different types of data. - EmbeddingGemma 2 sets a new standard for models under a billion parameters, even outperforming much larger models on several benchmarks. To make this run efficiently on edge hardware, it comes in a modular form factor with a maximum of 740 million parameters. It outputs 768 dimension vectors, but thanks to
01:05 Matryoshka representation learning, you can truncate outputs down to as low as 128 dimensions. With an 8,000 token context window, you can embed long documents, code bases, or extended voice recordings right on device for search and retrieval later. - Let’s see what this looks like in practice with the Google AI Edge Gallery app. First, we’ll look at Instant Media Search, where you can use natural language queries to directly retrieve photos and visual assets. I can also search within videos using Video Moment Finder.
01:38 Say I want to find a specific part of a birthday video. I can search for blowing out candles, and it will find that exact moment immediately. This all happens directly on my phone without any intermediate transcription, and no external API calls are required. So my entire search can happen completely offline. Now let’s see how these on-device capabilities can power even more complex pipelines. - When paired with lightweight generative models like Gemma 4, you can build RAG pipelines
02:08 and agentic workflows where sensitive user data never leaves the hardware. So let’s see an example of this in the AI Edge Foresight app. So I might say, “Hey, how does Foresight actually leverage EmbeddingGemma 2?” And Foresight’s actively listening in the background, it embeds the audio stream, detects the question, and instantly it surfaces a generated summary alongside the system architecture diagram retrieved straight from local drive storage. Because everything is embedded completely on device, sensitive information stays strictly private.
02:39 So while the model delivers incredible out-of-the-box quality, you can also fine tune this model for domain-specific vocabularies and specialized assets, whether that’s indexing complex legal contracts, medical imaging, or technical product catalogs. This helps improve retrieval precision without increasing model size. - Download the model weights today on Hugging Face and explore our notebooks in the Gemma Cookbook to get started.