Google DeepMind Launches EmbeddingGemma 2 for On-Device Multimodal Search

Google DeepMind has released EmbeddingGemma 2, an open-source model that turns text, code, images, video and audio into a single searchable format. It is built on the Gemma 4 architecture and licensed under the permissive Apache 2.0 license, meaning anyone can use it for commercial projects ¹.

The model has 740 million parameters and can run on consumer devices. Google reports it uses about 191 megabytes of RAM for text-only work and 567 MB when using all modalities on a Pixel 11 Pro ¹. That makes it one of the smallest multimodal embedding models available today.

One Space for Everything

An embedding model converts content like a document or photo into a list of numbers called a vector. Items with similar meaning end up close together, which makes them easy to search and compare. In a Retrieval Augmented Generation or RAG pipeline, these vectors let a large language model find fresh information it was not trained on ¹.

EmbeddingGemma 2 maps all five content types into one shared 768-dimensional space. A text query can retrieve an image. A voice memo can find a video clip ³. The architecture is modular: developers can load only the parts they need, scaling from 270 million parameters for text and code up to 740 million for the full model ².

Speed on Small Devices

Google designed the model for on-device use. Using a technique called quantization-aware training, weights are compressed to save memory without losing much accuracy. The company measured inference at 37.3 milliseconds per image on a MacBook M5 Pro GPU ¹.

The model also supports Matryoshka Representation Learning, or MRL, which lets developers shrink output vectors from 768 dimensions down to 512, 256 or 128 ³. Truncating to 128 dimensions can cut storage by up to six times, though Google recommends keeping 256 dimensions for multimodal work ¹.

The context window holds 8,192 tokens — four times larger than version 1 — enough for roughly 29 images, 58 video frames or 5.5 minutes of audio ¹.

Better at Code

On benchmark tests, EmbeddingGemma 2 scores leading results among models with fewer than one billion parameters. It earned 78.68 on MTEB Code, jumping nearly 10 points over version 1's 68.76 ¹. Multilingual text performance stayed strong at 61.36 on MTEB multilingual v2 .

Larger models still beat it on some tasks. Qwen3-VL-Embedding-2B, for example, reports higher scores but requires about 2.7 times more parameters and has no audio support ¹.

Weights are already live on Hugging Face and Kaggle, with builds available through Ollama, llama.cpp and LiteRT ¹. ML Kit support for Android with NPU acceleration is coming within weeks .

Key facts

EmbeddingGemma 2 uses 740M parameters and runs in under 567MB RAM on a phone
It maps text, code, images, video and audio into a single shared vector space
Code retrieval accuracy jumped nearly 10 points compared to version 1
Matryoshka Representation Learning lets vectors shrink to 128 dimensions, cutting storage six-fold
Released under Apache 2.0, weights are on Hugging Face and Kaggle