Open on-device model to search documents, images and audio together
In brief
Google has released EmbeddingGemma 2, built on Gemma 4, as open weights under the Apache 2.0 license. According to Google, it runs on laptops and phones for search, RAG, classification and clustering. It encodes text, code, images, video and audio into the same 768-dimension space, scaling from 270 million parameters for text and code to 740 million with all media types.
Source : InfoQ - « EmbeddingGemma 2 Expands beyond Text to Support Code, Images, Video, and Audio »
What it changes for an SME
A typical SME does not need to install this model. If you already operate internal search or RAG, Google says you could index documents, screenshots, recordings, code and video on a laptop or phone and compare them in one space, without a separate system per media type. The model is built to run on your own hardware and, according to Google, memory can be cut if you accept shorter vectors and some loss of quality, especially for images, video and speech. Until a tool you already use adopts it, day-to-day time, cost and risk for most firms of 10 to 250 people stay the same.
What you can do
- Ask the team or vendor that runs your internal search or assistants whether embeddings are computed on your machines or in a cloud API.
- If you already test local RAG, have a developer review EmbeddingGemma 2 on Hugging Face and Google's figures on quality when vectors are shortened.
- Do not assume text-search quality will match image, video or audio retrieval if you reduce vector size; Google reports a steeper drop for those media.
Let's talk about the time you could save.
In 20 minutes, we find where AI could save you time or money. Free, no commitment.