Google Releases EmbeddingGemma 2 for Offline AI Search Across Text, Images, Audio and Video

Google DeepMind has introduced EmbeddingGemma 2, a new artificial intelligence model designed to make searching across different types of digital content easier and more efficient. Announced on October 6, 2026, the model can process text, computer code, images, video and audio within a unified embedding space. This means information from different formats can be compared according to its meaning rather than relying only on exact words or filenames. For example, a user could search for a particular video clip using a text description or locate a recording by referring to something mentioned in the audio. Google says the model is designed for on-device use, reducing the need to send personal files to remote servers for processing.

One of the model’s main advantages is its relatively small size. EmbeddingGemma 2 contains 740 million parameters and is built on Google’s Gemma 4 architecture. Its modular design allows developers to use a smaller version for text and code or add specialised components for image, video and audio processing. According to Google, the text-and-code configuration uses 270 million parameters, while the full multimodal configuration uses 740 million. This flexibility allows developers to choose a setup that matches their devices and application requirements. Google also reports that the model can operate with approximately 191 megabytes of active RAM for text-only use and 567 megabytes for the full multimodal configuration when quantised on a Pixel 11 Pro.

EmbeddingGemma 2 works by converting information into numerical representations known as embeddings. These representations capture relationships in meaning, allowing a search system to identify relevant content even when a user’s query does not contain the exact words found in a file. For instance, someone could search a large collection of recordings for a discussion about business expenses without remembering the precise wording used during the meeting. The same approach can help developers search code repositories, organise photographs, retrieve information from documents and build AI assistants that reference local files. Google says the model supports an 8,192-token context window and can produce embeddings in dimensions ranging from 768 down to 128, allowing developers to reduce storage requirements when appropriate.

Privacy is another important focus of the release. Because EmbeddingGemma 2 is designed to run directly on supported devices, applications can generate embeddings without automatically uploading the underlying content to a cloud service. This can reduce data exposure and improve response times, particularly when working with private recordings, personal photographs, business documents or sensitive information. Developers can also combine the model with a generative AI system such as Gemma 4 to create retrieval-augmented generation applications, commonly known as RAG. In these systems, one model retrieves relevant information while another uses that information to generate an answer. However, local processing does not automatically guarantee complete privacy; the security of an application also depends on its storage, permissions, data handling and other implementation choices.

Google says EmbeddingGemma 2 delivers significant improvements over its predecessor in certain retrieval tasks. On the MTEB Code benchmark, the company reports that the score increased from 68.76 for the original EmbeddingGemma to 78.68 for the new model. Google also says EmbeddingGemma 2 performs competitively against larger embedding models on selected text, vision and audio benchmarks. These results are promising, although benchmark performance does not necessarily translate into identical results across every language, device or real-world application. The release is available under the commercially permissive Apache 2.0 licence, making it accessible to developers interested in experimenting with the model and incorporating it into eligible commercial projects.

The launch reflects Google’s broader effort to bring more AI capabilities directly to phones, laptops and other devices instead of depending entirely on cloud-based computing. Compact models can help developers build search tools that work with personal media libraries, offline assistants and applications that need to retrieve information quickly. EmbeddingGemma 2 could also support software development tools that search code, as well as workplace applications that locate relevant details across documents, recordings and presentations. Its effectiveness will ultimately depend on how developers integrate it into practical products and how well those products perform on different hardware. Nevertheless, by combining several content formats within one relatively compact model, Google DeepMind is giving developers another option for building AI search systems that are faster, more flexible and potentially more privacy-conscious.