Google Launches EmbeddingGemma 2: A 740 Million Parameter Multimodal AI Model

Google has officially launched EmbeddingGemma 2, a new multimodal embedding model boasting 740 million parameters, on October 6, 2026. This model is designed to operate on existing hardware, allowing developers to leverage its capabilities without relying on remote data centers.

Key Features of EmbeddingGemma 2

• Multimodal Capabilities: EmbeddingGemma 2 can process text, images, video, and audio, placing all content into a shared 768-dimensional vector space. This allows for seamless connections between different types of data, such as matching text queries to images.
• Increased Context Window: The model features an 8K token context window, which is four times larger than its predecessor, enhancing its ability to handle complex queries.
• Modular Architecture: At its core is a 270 million parameter backbone for text and code, with optional encoders for vision and audio, bringing the total to 740 million parameters.
• Privacy Focus: EmbeddingGemma 2 is tailored for privacy-sensitive applications, allowing embeddings to be generated directly on devices, thus keeping personal data secure.
• Matryoshka Representation Learning: This feature enables efficient storage by packing essential information into smaller vectors, which can reduce embedding dimensions significantly.

Developers can access the model weights under an Apache 2.0 license, facilitating commercial use and integration with various frameworks. The original EmbeddingGemma model achieved over 20 million downloads, indicating strong interest and utility in the developer community.