News
EmbeddingGemma 2 Released: 740M Parameters, One Model to Search Text, Images, Audio, Video
Google DeepMind released EmbeddingGemma 2 on October 6 — a 740M-parameter open embedding model mapping text, code, images, audio, and video into one vector space. It runs offline on phones and computers; searching local files needs no internet.
What Was Released
On October 6, Google DeepMind announced EmbeddingGemma 2 on its official blog. It’s an embedding model that turns text, code, images, audio, and video into vectors in a single vector space. Once converted, cross-modal search becomes a distance computation: give it a piece of text, and it finds the closest-matching image or video clip.
The model has 740M parameters, built on the Gemma 4 architecture. Weights are public on Hugging Face and Kaggle under the Apache 2.0 license — commercial use is fine.
The previous EmbeddingGemma came out last year and handled text only. Google says it was downloaded over 20 million times, with developers using it for on-device search and privacy-focused RAG. This generation’s biggest change: one model swallowing five input types.
What It Can Do
Google gave concrete examples. Use a voice memo as input to find the matching video clip. Describe something in words to locate content in hours of recordings. Text, images, audio, and video mixed together, handled by this one model.
It shares an audio encoder and tokenizer with Gemma 4. The two models can run together with lower total memory footprint. Paired with Gemma 4, developers can build fully offline local file Q&A: data never leaves the device, everything computes locally.
How Heavy Is It
Memory is what matters most for on-device models. Per reports, on a Pixel 11 Pro the text-only version uses about 191 MB of active memory, the full multimodal version about 567 MB. At that scale, ordinary phones run it directly.
On performance, Google calls it “the best open native multimodal embedding model in its class.” It also claims the model beats specialized models more than twice its size on several benchmarks. Reports add that its MTEB Code score rose from 68.76 to 78.68 — good for indexing local codebases.
Google’s Companion Moves
Alongside the model release, Google added new features to its on-device tooling.
AI Edge Gallery is Google’s app for developers to try models on-device. Two new features this time: Instant Media Search — search photos and videos on the phone with a natural-language description or a sample image; and Video Moments Finder — locate specific moments in local videos by text description, like “a kid laughing,” without transcribing audio first.
There’s also an experimental Mac app, Google AI Edge Foresight, for searching meeting notes, notes, images, and documents on the computer.
In the coming weeks, the model lands in Android’s ML Kit with support for hardware-accelerated devices. Google also wired it into the MediaPipe toolchain, so developers can build on-device search and classification on iOS, macOS, Windows, Linux, and Web.
Why It Matters
Two reasons. First, privacy. Search and RAG run entirely locally — files never upload to the cloud. For people handling sensitive content like meeting notes and codebases, that need is real.
Second, the license. Apache 2.0 means direct commercial use. 740M parameters and a few hundred MB of memory — independent developers can play too.
Generative models already exist (Gemma 4, Gemma 3n); now the retrieval piece is filled in. Next, let’s see what developers build.