Google Unveils EmbeddingGemma 2, Which Finds Video and Audio With No Internet

Google DeepMind unveiled EmbeddingGemma 2 on October 6 local time. The model converts text, code, images, video, and audio into the kind of numerical representations that search systems can work with. Its defining feature is the ability to link material across different formats. It was designed so that you can locate a specific moment in a video using a voice memo, or type a text query to find the passage you want inside a long recording.

What changed beyond text search

The scope of search has expanded from text alone to images, video, and audio. The previous model, EmbeddingGemma, focused on organizing and searching text on personal devices. According to Google DeepMind, that earlier release, introduced last year, has been downloaded more than 20 million times. The new model handles multiple formats with a single model, so a search works even when the format of your query differs from the format of the file.

Code search performance has also improved. Google DeepMind reports that its score on MTEB Code, a code performance benchmark, rose 9.92 points, from 68.76 with the previous model to 78.68. The company says performance on multilingual text remains on par with its predecessor. Developers can use it to find the snippet they need in code stored on a device, or to select the code that a coding assistant should reference.

Is it light enough to run on a phone?

EmbeddingGemma 2 is built on Gemma 4. The model comes in at 740 million parameters. Because it is structured so that you can activate only the capabilities you need, it shrinks to 270 million when handling text alone. Google DeepMind says it applied model compression techniques. Running on a Pixel 11 Pro, the memory required for the model itself is roughly 191MB for the text-only configuration and roughly 567MB with all capabilities enabled, according to the company.

Image: Google

The amount it can take in at once is four times that of the previous model. The company says it can process up to 5.5 minutes of audio, 29 images, or 58 video frames. Mixing formats in a single input also works. There is a feature that shortens the numerical representations used for search, which can cut the storage space and memory needed to keep them by as much as a factor of six.

Finding your own files without an internet connection

The core use case is search that begins and ends on the device. Google DeepMind says that generating the search representations directly on the device helps protect personal information. It adds that processing delays are reduced and that search works even without an internet connection. A developer building an app that surfaces private recordings or videos can process those files without sending them anywhere.

Image: Google

The model can be downloaded from Hugging Face and Kaggle. It is released under the Apache 2.0 license, which permits commercial use. Development tools for embedding it in mobile apps and browsers are also supported. Google DeepMind published guidance for developers who want to further train the model for their own purposes. The model will also be added to the Gemini Enterprise Agent Platform Model Garden at a later date.

Paired with generative AI such as Gemma 4, it can power features that search local material on a device and reference it when answering. How useful this proves for users in Korea will depend on the accuracy of Korean-language search and on how fast it runs on the devices people actually carry.