EmbeddingGemma 2: The Developer Guide- Google Developers Blog

Introduces EmbeddingGemma 2, a sub-1B multimodal embedding model based on Gemma 4, capable of processing text, code, images, video, and audio into a unified 768-dimensional space. Highlights its modular architecture allowing selective loading of encoders (270M text/code, 440M text+vision, 570M text+audio, 740M full multimodal) and Matryoshka Representation Learning (MRL) for dynamic dimension truncation (down to 128 dimensions) to reduce storage requirements. Provides developer guide for using the model with sentence-transformers library, including task prompts for retrieval and handling interleaved inputs. Benchmarks show improved performance over EmbeddingGemma 1 on code retrieval and multimodal capabilities.