Gemini Embedding 2: A Native Multimodal Embedding Model from Gemini

Google introduced Gemini Embedding 2, a native multimodal embedding model that learns shared representations for video, audio, image, and text.

Built with large-scale contrastive learning, it aims to improve cross-modal retrieval and search across different media types.

The model reports state-of-the-art results on several retrieval benchmarks, along with strong zero-shot performance in specialized domains.

It could be useful for search, recommendation, and RAG systems by providing a more unified general-purpose embedding space.