EmbeddingGemma 2: An open, lightweight multimodal embedding model (blog.google)

Google has launched EmbeddingGemma 2, an open-source, lightweight multimodal embedding model that enhances capabilities beyond text to include code, images, video, and audio. This model, built on the Gemma 4 architecture and featuring 740 million parameters, enables efficient on-device inference, making it possible to perform complex tasks such as finding video clips from audio or searching through audio recordings using text queries. The model's highlights include its best-in-class performance for sub-1B multimodal embedders, significant improvements in code performance, modular design for different workloads, and storage efficiency through dynamic truncation of output vectors. The significance of EmbeddingGemma 2 for the AI/ML community lies in its ability to deliver robust, privacy-friendly search and retrieval capabilities directly on edge devices. With features like an 8K token context window and compatibility with generative models for retrieval-augmented generation (RAG) workflows, this new model sets a new standard for quality and efficiency in the multitasking landscape. Its on-device nature reduces latency and enhances data privacy, positioning it as a vital tool for developers aiming to build advanced, efficient applications across various modalities.