Google Releases EmbeddingGemma 2, a Lightweight Multimodal Model That Runs On Devices

At a glance

Google DeepMind has released EmbeddingGemma 2, a lightweight AI model that handles text, images, audio, and video together. It can run embedding directly on devices such as smartphones. Embedding means turning data into a list of numbers so that similar items are easier to find. The model is released under the Apache 2.0 license.

Key points

1. 1

   The previous version only handled text, but this version can group and search code, images, video, and sound in one shared space.

2. 2

   It has about 740 million parameters (the learned values in an AI model), and it is designed to run well on devices.

3. 3

   With text alone, it works with only 270 million parameters, and image and sound features can be added separately.

4. 4

   The company says it can be used to find a specific video scene from a voice memo, or to search several hours of recordings by text.

Why it matters

If AI search and answer features can run on a user's own device, personal files do not have to be sent to outside servers, which can help protect privacy.