Google EmbeddingGemma 2 (twitter.com)

Google has announced the release of EmbeddingGemma 2, an advanced lightweight multimodal embedding model that integrates various data types—text, code, images, video, and audio—into a cohesive embedding space. This model is optimized for on-device applications, featuring a compact 740 million parameter architecture with modular encoders. Notably, it supports flexible dimensional sizes ranging from 768 to 128 dimensions through a technique called Matryoshka Representation Learning (MRL), significantly expanding its versatility for developers. The significance of EmbeddingGemma 2 lies in its larger 8K context window, which is four times the capacity of its predecessor's text-only version. This enhancement allows for richer, more nuanced understanding and processing of multimodal inputs. Additionally, the model comes under a commercially permissive Apache 2.0 license, making it accessible for a wide array of applications across the AI/ML community, from content generation to advanced analytics, further fostering innovation in the field.