SDSignal Desk

EmbeddingGemma 2 fits multimodal search of photos, audio and video into a phone’s memory

· Updated · Signal Desk newsroom

Analysis of reporting by Google DeepMind. How we work

Image: Google DeepMind

Google DeepMind’s EmbeddingGemma 2 is a 740 million-parameter open model under Apache 2.0 that embeds text, code, images, audio and video in one space. Google says it runs in about 567MB on a Pixel 11 Pro, making private, offline search of personal media practical, with new questions about who controls the device.

Key takeaways

  • EmbeddingGemma 2 has 740 million parameters and an Apache 2.0 license.
  • It maps text, code, images, audio and video into one shared embedding space.
  • Google says the quantized multimodal model needs about 567MB on a Pixel 11 Pro, about 191MB text-only.
  • Context grows to 8K tokens, four times the first version.
  • Google reports MTEB Code scores rising from 68.76 to 78.68.

What Google released

Google DeepMind released EmbeddingGemma 2, a 740 million-parameter open model under the Apache 2.0 license that maps text, code, images, audio and video into a single shared embedding space (release details). Embeddings let software find things by meaning rather than exact words, so they underpin search, retrieval and recommendation.

The notable change is where that plumbing can run. Google says that with quantization the full multimodal model needs about 567MB of active memory on a Pixel 11 Pro, and about 191MB for text alone. The context window is 8K tokens, four times the first version, which Google says covers roughly five and a half minutes of audio or 58 video frames. The first EmbeddingGemma passed 20 million downloads. The model is built on the Gemma 4 architecture and shares its text tokenizer and audio encoder, so the two can run together on one device with a smaller combined footprint.

Why on-device matters

Searching hours of recordings with a text query, or finding a video clip from a voice memo, has generally meant sending media to a server. A model this small running offline keeps that data on the device, in line with other local-first tools such as Google’s offline note-taker (offline transcription). The design is modular: a text-only app can use as little as 270 million parameters, with vision and audio encoders added only when needed.

Google reports code-retrieval scores on MTEB Code rising from 68.76 to 78.68, which makes local codebase search and retrieval for coding agents more realistic. Vector sizes can be trimmed from 768 dimensions to as few as 128, cutting storage up to sixfold. These are Google’s figures; independent tests on messy real-world media would show how well they hold. Weights are on Hugging Face and Kaggle, with support in transformers, llama.cpp, Ollama and vLLM, and Google says enterprise Model Garden availability is coming soon. The release joins a run of small open models aimed at local hardware (Nemotron 3 Diarization).

The other side of private search

Cheap, private, on-device search across every photo, recording and video is also cheap search for whoever controls the device. Stalkerware or anyone with access to a phone could sift a person’s media by meaning in seconds, with no server logs. Privacy from the cloud is not the same as privacy from the person next to you. As the capability becomes standard in apps, platform rules about which apps can index photos, audio and video will matter.

Overall the release pushes AI search toward the edge and puts responsibility on app makers to use it well. More in our Google Gemini coverage, Google coverage, tools coverage and research coverage.

Who feels it

App developers
A small, permissively licensed model for offline multimodal retrieval, with day-one tooling.
Phone and laptop users
Smarter search of personal media without uploading it, if apps adopt it.
Privacy and safety advocates
Less cloud exposure, but deep searches of a seized or shared device become easier.

What to watch

  1. Independent benchmarks on real personal media.
  2. Consumer apps that ship on-device search built on it.
  3. Model Garden availability for enterprises.
  4. Platform controls on which apps can index media.

FAQ

What is EmbeddingGemma 2?

A 740 million-parameter open multimodal embedding model from Google DeepMind under Apache 2.0.

Which inputs does it handle?

Text, code, images, audio and video in one embedding space.

How much memory does it need?

About 567MB for the full quantized model on a Pixel 11 Pro and about 191MB text-only, Google says.

How long is its context?

8K tokens, roughly five and a half minutes of audio or 58 video frames.

Where can developers get it?

Hugging Face and Kaggle, with support in transformers, llama.cpp, Ollama and vLLM.

Read the original

Continue at the source.

Google DeepMind

More on AI

All AI coverage →