Google DeepMind launches EmbeddingGemma 2, a 740M-parameter model to map code, images, video, and audio in a shared embedding space, under an Apache 2.0 license

On-device multimodal AI search capabilities become accessible for developers using a single, lightweight model.

Google DeepMind has released EmbeddingGemma 2, an open-source, multimodal embedding model with 740 million parameters. This model is designed for on-device inference and can natively map combinations of text, code, images, audio, and video into a unified embedding space. Based on the Gemma 4 architecture and released under the Apache 2.0 license, EmbeddingGemma 2 offers modularity, allowing developers to use only text encoders (270M parameters) or add vision (170M) and audio (300M) components. It features Matryoshka Representation Learning for efficient storage and an 8K token context window. The model achieves strong performance across text, code, image, and audio benchmarks, outperforming some larger models. It is optimized to run on devices like the Google Pixel 11 Pro, requiring minimal RAM and storage.

EmbeddingGemma 2's release signifies a significant step towards democratizing advanced on-device AI processing, particularly for multimodal applications. By offering a highly capable, yet compact model under a permissive license, Google DeepMind is enabling a wider range of developers to build sophisticated AI features directly into consumer hardware. This move challenges the dominance of cloud-based AI services for certain tasks by providing a privacy-preserving, low-latency alternative that functions offline.

The model's modular design and storage-efficient features like Matryoshka Representation Learning are key differentiators, allowing for tailored deployments that balance functionality with resource constraints. Its ability to handle diverse data types like code, images, and audio within a single embedding space opens new avenues for semantic search, content retrieval, and intelligent agents operating directly on edge devices.