Google Releases EmbeddingGemma 2: Five Modalities in a 768-Dimensional Space

On October 6, 2026, Google DeepMind released EmbeddingGemma 2, an open embedding model for text, code, images, video and audio. It projects all five data types into a single 768-dimensional space. According to Google, on a Pixel 11 Pro the text part takes about 191 MB of RAM, and the full multimodal version about 567 MB. That is the footprint of a mid-sized app, not of server infrastructure.

What the model can do

An embedding turns an object into a vector of numbers, and objects with similar meaning end up close together. Until now, developers have typically kept separate models for different data: one for text search, another for images, a third for audio. Their vectors could not be compared with each other, because each model has its own coordinate system.

EmbeddingGemma 2 has one shared coordinate system. The query "a dog catches a frisbee in the park" produces a vector that lies close to a video clip on that topic, to a photograph and to a voice note. For a developer, that means one index instead of three and one query instead of three.

The first EmbeddingGemma came out in September 2025. It was a text model of roughly 300M parameters that fit into 200 MB of memory after quantization. The second version keeps the 768 dimensionality and adds the other modalities. For the text part, the memory budget stays in the same class.

Key specs

According to Google and a breakdown on Habr:

• release: October 6, 2026, developed by Google DeepMind;
• type: open embedding model;
• modalities: text, code, images, video, audio;
• dimensionality of the shared space: 768;
• memory on Pixel 11 Pro: about 191 MB for the text part;
• memory on Pixel 11 Pro: about 567 MB for the multimodal version;
• use case: on-device search and data matching, without sending data to a server.

The available summary does not include a full set of benchmarks, the parameter count or the license terms. Without them, it is too early to compare the model with competitors. The memory figures are given for a specific smartphone and may differ on other hardware.

Why local execution matters

The key point of the release is where the model runs. Embeddings power search, recommendations and RAG, which means they sit where AI touches personal data: photos, recorded conversations, messages, home video. When vectorization happens on the device, those files do not leave the phone. That resolves some privacy concerns, removes network latency and makes it possible to search without an internet connection.

The memory gap is telling, too. The multimodal version is almost three times heavier than the text one, but 567 MB still fits within the budget of a modern smartphone. Video and audio cost more to compute: the encoder has to process frames and spectrograms rather than short strings.

What to check before adopting it

The author of the Habr breakdown promises to show the pitfalls. Several things are worth checking in advance:

• cross-modal search quality on your own data, not just on public benchmarks: a shared space is, by definition, a compromise between modalities;
• video and audio indexing speed on your target devices, because that will be the bottleneck on the first pass through a gallery;
• the license for the open weights, if you plan commercial use;
• vector storage size: at 768 dimensions, the index of a large media library will grow noticeably.

For teams building search over personal data or offline assistants, EmbeddingGemma 2 looks like a workable foundation for a prototype. It is best to start with the 191 MB text version and add the multimodal one once it is clear that video and audio are needed in the product.