Google puts multimodal search directly on your phone

EmbeddingGemma 2 is a lightweight, open-source embedding model built to bring cross-modal search straight to consumer devices. Released under an Apache 2.0 license, the full model packs 740 million parameters and maps text, code, images, audio, and video into a single shared vector space.

Unlike traditional generative models, it is not designed to write text or chat. Instead, its job is understanding and indexing. You can use it to search through hours of audio recordings with a text query, or use a short voice memo to pinpoint an exact video clip, all processed locally without sending personal data to the cloud.

How it works under the hood

Building local RAG pipelines or mobile search usually meant stitching together separate, memory-heavy models for every data type. Built on the Gemma 4 architecture, EmbeddingGemma 2 solves this by handling multimodal data natively in one model.

It features a clever modular design. If your application only indexes text or code, you can run a slim 270M parameter base. Adding vision requires another 170M parameters, and audio adds 300M. The entire multimodal package takes up only around 1.5 GB on disk.

To keep storage footprints practical on mobile hardware, the model supports Matryoshka Representation Learning (MRL). Developers can dynamically truncate the standard 768-dimensional output vectors down to 512, 256, or 128 dimensions. This can shrink local vector database storage and memory usage by up to 6x.

Specs, performance, and hardware

The model delivers standout performance for its size. On the MTEB Code benchmark, it jumped from 68.76 to 78.68, an improvement of nearly 10 points that makes it especially capable for local codebases and coding agent retrieval. Google notes it also matches or outperforms specialist models more than twice its size on audio benchmarks like MAEB and image evaluations.

Its context window has been expanded to 8K tokens, which is four times larger than the original EmbeddingGemma. That capacity allows it to digest up to 5.5 minutes of audio, 29 images, 58 video frames, or interleaved multimodal inputs directly on edge hardware.

Hardware requirements are minimal. After quantization on a Google Pixel 11 Pro, text-only inference requires just ~191MB of active RAM, while running the entire multimodal pipeline consumes only ~567MB.

How to run it

The model weights are live on Hugging Face and Kaggle under the permissive Apache 2.0 license.

Ecosystem support is ready out of the gate. For local inference and serving, it works with Ollama, llama.cpp, LMStudio, MLX, vLLM, transformers, and sentence-transformers, with vector storage support in Qdrant. Unsloth has also provided guides for custom fine-tuning.

For edge and on-device builds, developers can deploy via Google AI Edge MediaPipe and LiteRT, or target web browsers using transformers.js and WebGPU.

Why it matters

For developers building local-first applications, personal assistants, or offline RAG pipelines, EmbeddingGemma 2 is one of the most practical releases this year. Consolidating text, audio, and visual search into an under-600MB footprint removes the infrastructure headache of juggling multiple heavy models on consumer devices.

We think Google made a smart move by modularizing the encoders and aligning tokenizers with Gemma 4, which significantly reduces overhead when running retrieval and generation together. Keep in mind that while an 8K window is generous for mobile tasks, longer audio recordings or feature-length video will still require careful client-side chunking.