Google releases EmbeddingGemma 2, a sub-1B open multimodal embedding model

On October 10, 2026, Google announced EmbeddingGemma 2 on the Google Developers Blog. It is a sub-1B open-source multimodal embedding model built on Gemma 4 and released under the Apache 2.0 license. According to the post, it maps text, code, images, video, and audio into a unified 768-dimensional space.

For developers, the model offers modular modality encoders that can be loaded selectively through the sentence-transformers library. These encoders range from 270M to 740M parameters, letting teams tune memory usage to their needs. EmbeddingGemma 2 also applies Matryoshka Representation Learning, which supports dynamic dimension truncation down to 128d. Google says this can significantly reduce vector database storage requirements.

The announcement frames the release around a common problem: modern search and retrieval augmented generation (RAG) applications increasingly need to handle diverse content types, from technical documentation and source code to images, video clips, and audio recordings. The difficulty, according to the post, is finding models that deliver strong retrieval accuracy while remaining practical to run. The report does not go beyond the launch details, so the public information currently stops at the model's availability, its licensing terms, and its stated loading, parameter, and dimension-truncation capabilities.