EmbeddingGemma 2: An open, lightweight multimodal embedding model

On-device multimodal embedding models are now available that can process text, images, audio, and video, enabling offline, private search across different media types.

Google DeepMind has launched EmbeddingGemma 2, an open-source, lightweight multimodal embedding model designed for on-device inference. This successor to EmbeddingGemma expands beyond text to natively map combinations of text, images, audio, and video into a unified embedding space. Built on the Gemma 4 architecture, it features 740 million parameters and is released under an Apache 2.0 license, making it suitable for commercial use. The model achieves state-of-the-art scores for its size across benchmarks like MTEB Code and MAEB, while offering modularity with optional encoders for vision and audio. It supports dynamic vector truncation for storage efficiency and includes an 8K token context window for processing longer sequences. The goal is to enable privacy-preserving, low-latency, offline multimodal search and retrieval directly on consumer hardware.

EmbeddingGemma 2's modular design allows developers to tailor its parameter count and functionality, from text-only workloads using just 270 million parameters to full multimodal support with additional encoders for vision and audio. This flexibility, combined with Matryoshka Representation Learning for dynamic output vector truncation, offers significant storage and memory savings, making it ideal for resource-constrained edge devices. The model's extended 8K token context window further enhances its capability for processing richer, longer multimodal inputs directly on hardware.

The release of EmbeddingGemma 2 signals a significant step towards more capable and accessible on-device AI applications, particularly for privacy-sensitive use cases like retrieval-augmented generation (RAG) pipelines that operate entirely offline. Its native multimodal embedding capabilities and compatibility with other Gemma models promise to simplify the development of integrated AI systems for tasks ranging from local codebase indexing to real-time decision engines, fostering innovation in consumer electronics and edge computing.