TL;DR: Google DeepMind launched EmbeddingGemma 2 on Oct. 6, 2026, a 740-million-parameter open model that maps text, code, images, video and audio into one on-device embedding space.

Google has moved its on-device search model beyond words. On Oct. 6, 2026, Google DeepMind introduced EmbeddingGemma 2, an open-weight embedding system that places text, code, pictures, video frames and sound in a shared numerical space for phones and other local hardware, according to Google's announcement on its Keyword blog. The Verge reported the same launch that afternoon, noting the 740-million-parameter size and examples such as locating a video segment from a spoken note or scanning lengthy audio with a typed query.

The release shifts local retrieval from single-mode lookup to cross-mode lookup. Instead of chaining a caption writer, a speech transcriber and a text embedder, one compact model converts varied inputs into comparable vectors. Those vectors can then be stored on the device, compared for similarity, and used for offline search, routing and retrieval-augmented generation. Google positions the approach as privacy-preserving and faster because private files do not need to leave the handset for ranking.

What EmbeddingGemma 2 does

EmbeddingGemma 2 is described by Google as its most capable system for multimodal embeddings on local devices. It natively handles combinations of written language, program code, still pictures, moving images and sound, producing 768-dimension vectors that sit in the same space regardless of origin.

That shared space is the central idea. A typed request can match a photograph. A picture can retrieve related clips. A hummed or spoken description can surface a particular passage in a recording. Google's examples include pulling an exact clip from a voice memo and hunting across hours of recordings with a short prompt, a use case highlighted by The Verge in its Oct. 6 coverage.

The model builds on the Gemma 4 architecture and is distributed under the Apache 2.0 license, which permits commercial use. Google said it draws on the same underlying technology as its Gemini Embedding family, but is sized and packaged for consumer hardware rather than data-center inference.

It is also intended to work alongside generative systems. When combined with Gemma 4, EmbeddingGemma 2 can supply retrieved files and passages for local retrieval-augmented generation. Because the two share a text tokenizer and an audio encoder, Google said running them together lowers the combined memory requirement compared with pairing unrelated models.

Why Google built a multimodal follow-up

The predecessor, EmbeddingGemma, arrived last year as a text-only option for organizing, searching and linking information directly on consumer machines. Google said developer uptake exceeded expectations, with more than 20 million downloads powering local search tools and private retrieval pipelines.

EmbeddingGemma 2 extends that base to other signals that live on phones: camera rolls, downloads, voice notes, meeting recordings and repositories. The timing reflects how much personal material is now multimodal. People no longer store only documents; they store screenshots, short videos, podcasts, lectures and code folders that need the same kind of semantic lookup.

The Google Developers Blog, in a companion Oct. 6 post from the Google AI Edge Team and Android ML Team, framed the launch as a way to cut delay and memory cost for builders. Rather than operating separate vision and speech front ends before embedding, developers can call one model to index media and answer descriptive queries locally, including zero-shot routing where an input is matched directly against labels and descriptions without additional training.

Inside the 740-million-parameter design

The total size is 740 million parameters, but the system does not have to load as a single block. Google structured it as a core plus optional encoders.

The text and code core accounts for about 270 million parameters. A vision encoder adds about 170 million, and an audio encoder adds about 300 million. A developer building only document search can deploy the text portion alone; one building full photo, video and sound search can add the other modules. Vectors from different configurations remain comparable, so a query encoded with the text-only setup can still match items indexed with the full multimodal setup, according to Google's developer guidance.

Memory figures center on a recent Pixel phone. With quantization on a Pixel 11 Pro, Google reported roughly 191MB of active RAM for text-only weights and about 567MB for the complete multimodal package. That envelope is meant to fit within the tight budgets of current handsets and laptops while leaving room for an app, an index and a generative partner model.

Storage for the index itself can shrink as well. Using Matryoshka Representation Learning, developers can truncate output vectors from 768 dimensions to 512, 256 or 128 dimensions. Google said that step can deliver up to a sixfold reduction in local vector-database size and memory use, a relevant saving for large camera rolls or audio archives.

Context capacity has grown. The model offers an 8,000-token context window, four times the size of the first EmbeddingGemma. In practical terms, Google said a single input can cover up to 5.5 minutes of audio, 29 images, 58 video frames, or interleaved mixes of those forms on local hardware.

How it scores on code, text, vision and audio

For its size class, Google calls EmbeddingGemma 2 best-in-class among sub-1-billion-parameter multimodal embedders. The company cited leading results on MTEB Code for programming-related retrieval and MAEB for audio embedding, while matching or beating many larger systems on text, vision and audio assignments.

The clearest numeric change is in code. Google reported a rise of 9.92 points on MTEB Code, from 68.76 to 78.68, preserving strong multilingual text behavior while improving local codebase indexing, semantic code search and retrieval for coding assistants.

On images, video, documents and sound, Google said the model sets a new quality-per-parameter mark for systems under one billion parameters and even surpasses some specialist models more than twice its size. Full evaluation tables and model details are published in the EmbeddingGemma 2 model card referenced in the Keyword announcement.

The Google AI Edge post adds a latency claim for a different role: decision engine. It said the model can match user inputs against classification labels and descriptions in milliseconds, enabling instant intent routing without training data or fine-tuning. That positions embeddings not only for search, but for on-device classification, triggering and predictive routing through interfaces such as the MediaPipe Decision Task API.

Apps that show offline media search

Google is demonstrating the model through two additions to its Google AI Edge Gallery showcase app and a Mac experiment.

Instant Media Search lets people find specific images or videos in local storage with everyday language or an example picture. Both the query and the media are turned into on-device vectors, with media vectors held in a local SQLite database. Retrieval ranks items by cosine similarity, so results refresh as the user types. The developer post illustrates multilingual behavior: typing the German word for cat immediately surfaces feline photos, while extending the phrase to describe a cat sleeping on a keyboard re-ranks the grid to the intended shot. Users can also start from a phone image or a live camera frame to pull visually and semantically related photos, with embedding and ranking handled through MediaPipe UniversalEmbedder and SemanticRetriever.

Video Moments Finder applies the same cross-modal comparison to moving pictures. After a person selects a local recording, such as a family video or sports reel, the on-device engine indexes visual keyframes alongside audio chunks. A descriptive prompt — the developer examples include kids laughing, a dog catching a frisbee and a person blowing out birthday candles — is embedded and compared against frame embeddings to highlight matching timestamps, without first transcribing speech or generating captions.

A third showcase is Google AI Edge Foresight for Mac, described as an experimental meeting companion. Powered by EmbeddingGemma 2 together with Gemma 4, it works with system audio and a microphone across meeting platforms, including when offline. Google listed enhanced note-taking that enriches shorthand with details pulled from the live conversation, plus cross-modal lookup across transcripts, notes, images and documents kept on the machine.

Google said the Gallery app can be downloaded from the Google Play Store and Apple App Store, with implementation code on GitHub for developers who want to study the embedding, storage and ranking flow.

Where developers can get it

Weights are available now on Hugging Face and Kaggle, with on-device-optimized variants through the LiteRT Community on Hugging Face. Google said availability in the Gemini Enterprise Agent Platform Model Garden will follow soon.

For building, Google points to MediaPipe for packaged embedding, retrieval and decision tasks and to LiteRT for custom integration, plus browser paths with transformers.js and WebGPU. The model also works with familiar tooling including Transformers, Sentence-Transformers, MLX, vLLM, llama.cpp, SGLang, Ollama and LMStudio, while vectors can be stored with systems such as Qdrant. Tuning guidance is provided by Unsloth.

On Android, Google said EmbeddingGemma 2 will become available as a service through ML Kit in the coming weeks, with NPU acceleration to broaden performance across devices. The company also published a Google AI Edge blog post on constructing on-device search and retrieval pipelines with LiteRT for teams ready to test private semantic search, visual keyframe retrieval and condition-triggered workflows.

Sources referenced