EmbeddingGemma 2 title graphic on a dark blue layered background — Google DeepMind released EmbeddingGemma 2 on October 6, 2026. Image: Google.

Semantic search means finding things by meaning, not by matching exact words. Until now, the heavy part of that work usually ran on someone else's cloud server.

On October 6, 2026, Google DeepMind released EmbeddingGemma 2. It is an open model built to do that work on a phone or laptop. It handles text, code, images, video and audio in one model, under the Apache 2.0 license.

I'm Nitin from Debugged Pro. We build search, chat and content apps for clients in India and the USA. "Can search work offline?" is a question we hear a lot. So I read Google's EmbeddingGemma 2 announcement, its developer guides and the full model card. Here is what is new, what the numbers say and where I would be careful.

What Google Released on October 6

An embedding model turns content into a list of numbers called a vector. Items with similar meaning get vectors that sit close together. Your app compares those vectors to find the best match.

The first EmbeddingGemma came out last year and handled text only. Google says it passed 20 million downloads. Version 2 adds images, video and audio, and it is built on the Gemma 4 architecture.

Here are the headline facts from Google's posts:

Size: 740 million parameters in total.
Modular parts: a 270M text and code core, plus a 170M vision encoder and a 300M audio encoder that load only when needed.
One shared space: every input becomes a 768-number vector, so a text query can match a photo or a sound clip.
Languages: support for more than 100 languages.
License: Apache 2.0, which allows commercial use.

Why one model for every format matters

Before this, a typical media search setup chained several models. One wrote captions for images, another turned speech into text, and a third embedded the text. Google's AI Edge team says EmbeddingGemma 2 cuts the delay and memory cost of that chain.

In simple terms, you can search a video by describing a moment, with no transcript or caption step in between.

Diagram showing text, code, image, video and audio inputs flowing through modular vision and audio encoders into EmbeddingGemma 2 and a shared embedding space — EmbeddingGemma 2 sends every input type into one shared vector space. Image: Google for Developers.

How Semantic Search Works With EmbeddingGemma 2

The flow is the same as any vector search system. The difference is that it can all happen on the user's device.

Index your content. The model turns each document, photo, clip or recording into a vector.
Store the vectors. Google's demo app keeps them in a local SQLite database.
Embed the query. The user types, speaks or picks an image, and that becomes a vector too.
Rank by similarity. The closest vectors come back as results.

Google showed this in its AI Edge Gallery app. In one demo, typing the German word for "cat" shows every cat photo on the phone. Typing more words re-ranks the grid as you go.

Load only what you need

This part matters most for app size and battery life. Per the developer guide, you can load four setups from the same checkpoint:

Setup	Parameters	Good for
Text and code only	270M	Docs, help centres, code search
Text plus vision	440M	Photo, PDF, slide and video search
Text plus audio	570M	Voice notes, call recordings, podcasts
Full model	740M	Mixed media libraries

All four share one vector space for semantic search. Start text-only and add images later, and you do not need to recompute old embeddings, Google says.

The Numbers Behind the Model

Google gives a lot of detail across its blog posts and the EmbeddingGemma 2 model card. These are the figures I would share with a client:

Memory on a phone: with quantization, about 191MB of active RAM for text-only weights and about 567MB for the full model on a Pixel 11 Pro.
Context window: 8,192 tokens, four times the first version. That fits about 29 images, 58 video frames or 5.5 minutes of audio in one input.
Code search: the MTEB Code score rose from 68.76 to 78.68.
Multilingual text: 61.36 on multilingual MTEB, roughly level with version 1 at 61.15.
Input rules: video is sampled at one frame per second by default, and audio should be 16 kHz mono.

Speed on real devices

Google's AI Edge post lists image embedding times on several devices. These used a budget of 70 vision tokens per image:

Device	Chip used	Time per image
MacBook M5 Pro	GPU	37.3 ms
Pixel 11 Pro XL	NPU	48.9 ms
iPhone 18 Pro	GPU	69.8 ms
iPhone 18 Pro	CPU	191 ms
Raspberry Pi 5	CPU	1,761 ms

The gap between the GPU and CPU rows is the real lesson for semantic search on phones. Indexing a large photo library on an older phone CPU will be slow, so plan to index in the background.

Chart of Massive Text Embedding Benchmark code scores against model size, with EmbeddingGemma 2 near the top for its size — Google's chart places EmbeddingGemma 2 high on code search for a model of its size. Image: Google for Developers.

Smaller vectors, smaller bills

The model supports a trick called Matryoshka Representation Learning. You can cut each vector from 768 numbers down to 512, 256 or 128. Google says this saves up to six times the storage.

The trade-off is clear in the developer guide. At 256 dimensions, text and code keep most of their quality, and image, video and speech keep about 95%. At 128, text keeps around 90%, but media search drops to around 75%.

Google's own example puts it in plain numbers. One million full vectors take about 1.5GB in bfloat16 precision. At 128 dimensions, they need just 250MB.

Limits Every Team Should Know

Google is upfront about weak spots, and The Next Web's report picked up on them too. Here is what I would flag before any build:

No safety tuning. The model card says it has no post-training alignment, safety tuning or output moderation. Filtering happened on the training data only.
Uneven languages. It supports 100+ languages, but Google says performance may not be equal across them.
Bias can carry over. The card warns that biases in training data can show up in the results.
Nuance is hard. Sarcasm and figurative language can trip it up.
Some pieces are not ready yet. Google says ML Kit support on Android will arrive in the coming weeks. Model Garden access is also listed as coming soon.

Since an embedding model does not write text, the risk sits in how you use its results. A search box that surfaces the wrong private file is still a real problem.

The India angle

Many Indian users search in Hinglish or mix English with a regional language. Google has not published a per-language breakdown for Indian languages in the material I read. Test with real queries from your users before you promise good results in Hindi, Tamil or Marathi.

On the plus side, on-device search keeps personal data on the phone. For apps that handle customer records, that can make planning for India's Digital Personal Data Protection Act, 2023 simpler. The same goes for US privacy rules.

Where On-Device Semantic Search Fits for Agencies

Most of our clients are not building photo apps. But many of them hold content that is hard to search today. Here is where I see this model fitting:

Field service apps: technicians search manuals and past job photos with no signal on site.
E-commerce: one vector can describe a product's text, photo and video together, so "waterproof trail shoes" finds the right listing.
Internal knowledge tools: search meeting recordings, PDFs and slides in plain language.
Developer tools: index a codebase locally for code search and coding agents.
Healthcare and finance: private retrieval where data should not leave the device.

It also pairs with Gemma 4 for on-device RAG, where a chatbot answers using your own files. Both models share a text tokenizer and audio encoder, so together they use less memory.

Comparing options for text-heavy work? Our breakdown of the Microsoft Harrier open source embedding model is a useful reference point.

Tools you can use today

Google lists support across most popular tools. The weights are on Hugging Face and Kaggle. You can serve the model with sentence-transformers (version 6.1.0 or later), transformers, vLLM, llama.cpp, SGLang, MLX, Ollama and LM Studio.

For mobile and web, Google points to MediaPipe, LiteRT, transformers.js and WebGPU. Qdrant is named for vector storage, and Unsloth for fine-tuning.

How to Pilot Semantic Search in Your App

If you want to test EmbeddingGemma 2 on a client project, this is the order I would follow:

Pick one narrow use case. For example, searching a help centre or a product catalogue.
Collect 200 to 500 real queries. Use search logs or support tickets, in the languages your users actually type.
Mark the right answers. Have two people agree on the correct result for each query.
Start with the smallest setup. Use the 270M text model unless you truly need images or audio.
Test vector sizes. Compare 768, 256 and 128 dimensions on your own data before you cut storage.
Use task prompts. Google trains the model with short prompts, so embed queries and documents with their matching prompt names.
Measure on real phones. Check speed and battery on a mid-range Android device, not only a flagship.
Add filters for private content. Make sure search only returns files the signed-in user is allowed to see.

Need help building a search or RAG feature like this? Our LLM development services cover model selection, testing and app integration.

Questions to ask before you choose an embedding model

Does it work well on our content and our users' languages?
Must search work offline, or is a cloud API fine?
How big will the index get, and can a phone hold it?
Will we need to search images, audio or video later?
Who can see the data, and where does it travel?

Key Takeaways

Google released EmbeddingGemma 2 on October 6, 2026. It is an open, Apache 2.0 model for semantic search across five content types.
It is small and modular, starting at 270M parameters for text and reaching 740M for every format.
It can run fully on device, needing about 191MB of RAM for text on a Pixel 11 Pro.
Smaller vectors save storage, up to six times, but media search quality drops at 128 dimensions.
It has no safety tuning and uneven language support, so test on your own data first.

Planning in-app search, a knowledge tool or an offline RAG feature? Our team offers free software development consulting for agencies and businesses in India and the USA.

About the Author

Nitin Kumar is the Lead Project Coordinator at Debugged Pro. He works with development, QA and delivery teams across India and the USA to keep client projects on track. He writes about the AI and software news that changes how agencies plan, build and ship work.