Google teaches EmbeddingGemma 2 to understand text, images, video, and audio

A multimodal model that fits in your pocket

On October 6, 2026, Google DeepMind released EmbeddingGemma 2, an open embedding model that addresses a core problem in the industry: previously, each data type required its own model, and the resulting vectors from different models couldn't be compared. A text vector couldn't be placed alongside an image vector. EmbeddingGemma 2 places all five modalities into a single 768-dimensional space.

The model processes text, code, images, video, and audio. This approach unlocks new capabilities: searching video clips using a text query, finding images within documents, or using a voice query to search a text database. On-device, this runs efficiently: on a Pixel 11 Pro, the text component takes up about 191 MB of RAM, while the full multimodal version takes up about 567 MB. That makes it possible to run this kind of search directly on the phone, without sending data to a server.

The first version of EmbeddingGemma launched in September 2025, supported only text processing, and according to the Japanese outlet livedoor, gathered more than 20 million downloads. The success of the first version pushed Google to develop a multimodal successor.

Architecture and technical specifications

EmbeddingGemma 2 is built on the Gemma 4 architecture. The model contains 740 million parameters, but it's not a monolithic structure — components are distributed modularly: 270M parameters for the text component, 170M for image processing, and 300M for audio. This architecture provides deployment flexibility: you can use just the text component or assemble the full multimodal version depending on available resources.

The model understands more than 100 languages and shows improvements over the first version: performance on code-related tasks is up roughly 14%. The context window has been expanded to 8192 tokens — four times larger than in the first version, allowing the model to process longer sequences without losing context.

A standout feature is Matryoshka embeddings: the model supports multiple output vector dimensions — 768, 512, 256, or 128. This allows developers to choose the optimal balance between accuracy and speed depending on the task. The text component is implemented as a compact transformer with 4 KV heads, 2 of which operate locally and 1 globally, providing efficiency without a significant increase in computational load.

Optimized for mobile devices

The model is specifically optimized to run on smartphones and laptops. That means semantic search, retrieval-augmented generation (RAG), data clustering, duplicate detection, and recommendation systems can all run entirely locally, without sending users' private data to cloud servers. Google has released the model openly as a set of weights and code, allowing the community to integrate it into various applications and services.

An embedding is a vector — a list of numbers that describe the meaning of content. Things that are similar in meaning get vectors that are close together; for example, the phrases "Today's wonderful weather" and "It's so sunny outside!" would end up near each other in vector space. Search works simply: a vector is computed for the query, vectors are computed for all documents, and the nearest neighbors are found using cosine similarity. This principle underlies semantic search, RAG (document search combined with generated answers), clustering, duplicate detection, and personalized recommendation systems.

The second version of EmbeddingGemma solves the problem of incompatible vector representations: all modalities — text, code, images, video, and audio — are converted into a single 768-dimensional space, making it possible to compare objects of different types within the same task.