Google has a new multimodal embedding model, and it wants it running on your phone. EmbeddingGemma 2 is a lightweight model that handles text, code, images, video, and audio in a single package, and the company is pitching it as a privacy-first tool for building search and retrieval apps directly on consumer hardware.
The model has 740 million parameters, runs under a commercially permissive Apache 2.0 license, and is built on the Gemma 4 architecture. It is meant to work within the tight memory and processing limits of edge devices — on a Google Pixel 11 Pro, the text-only weights take as little as ~191MB of active RAM, with the full multimodal model requiring ~567MB.
What EmbeddingGemma 2 Does
The model produces a shared embedding space for all five modalities. The idea is that a voice memo, a video clip, and a text query can all be compared against each other using the same underlying representation. That makes it possible to search through hours of audio based on a text query, or to find a specific video clip from a spoken description.
The design is modular. Developers can use the full 740 million parameters, or they can dial it back to 270M for text-only workloads. The optional vision and audio encoders come in at 170M and 300M parameters respectively. Storage efficiency comes from Matryoshka Representation Learning (MRL), which lets output vectors be truncated down to 768, 512, 256, or 128 dimensions — up to 6x reduction for local vector databases.
The Numbers Behind It
- Parameters: 740 million total, 270M for text-only, 170M vision, 300M audio
- RAM: ~191MB for text-only on Pixel 11 Pro, ~567MB for full model
- Context window: 8K tokens, up to 5.5 minutes of audio, 29 images, 58 video frames
- Code score: 9.92-point improvement in MTEB Code (68.76 to 78.68)
- License: Apache 2.0
Why It Matters
The model is being positioned as a foundational component for on-device RAG pipelines. Google says the model’s shared tokenizer means it can be paired with generative models like Gemma 4 in a unified pipeline with a lower combined memory footprint. That pairing enables retrieval augmented generation that works entirely offline, on the device itself, which is the whole pitch.
The company is also pushing deployment tools. MediaPipe handles embedding and retrieval tasks across platforms, while LiteRT lets developers integrate custom models. The browser side gets transformers.js and WebGPU. For serving, the list of supported libraries is long: transformers, sentence-transformers, MLX, vLLM, llama.cpp, SGLang, Ollama, and LMStudio. Storage is handled by Qdrant.
The Lightness Of It All
The model is being promoted as a lightweight alternative to larger multimodal embeddings. Google says it beats or matches many larger models across text, vision, and audio tasks in benchmarks like MTEB Code and MAEB. Code performance in particular saw a 9.92-point improvement in MTEB Code from 68.76 to 78.68.
The company is also tying this to a larger pattern. The announcement makes clear that embedding is moving toward smaller, faster, more efficient models that can run on the device itself rather than requiring a trip to the cloud. That is the direction of the market.
What We Make Of It
The announcement is a good example of how AI companies are moving toward on-device processing. The model is small enough to run on a phone, which means it can do multimodal search without sending data to the cloud — a privacy feature, not just a performance trick. That is the point Google is making, and it is a point worth making.
The modular design and the MRL truncation are real technical details, and they matter. But the pitch is simple: this is a model that can hear a voice, see a picture, and read a line of code, all at once, on a device you carry. That is the future of search, and Google is shipping it today.
The announcement is a good read. It is worth a look for anyone building on-device apps, and it shows how quickly these models are moving from research into everyday tools.
Source material: “EmbeddingGemma 2: An open, lightweight multimodal embedding model,” Google.
Get the Notebook.
The day's best stories and every fresh verdict, in plain English, in your inbox by seven. One email a day, no more.