Google Launches Gemini Embedding 2, Its First Multimodal Embedding Model

Most embedding pipelines require separate models per modality then fusion logic to compare across them. gemini-embedding-2-preview ingests interleaved inputs natively, collapsing that into one step and simplifying multimodal RAG, semantic search, and clustering. Dimensions flex from 128 to 3072 via Matryoshka Representation Learning — a nesting technique that lets developers tune performance vs. storage.

Use the Gemini API or Vertex AI to build multimodal search — query a video archive by text prompt, or use an image to retrieve matching content. Google reports outperformance over leading models on text, image, and video benchmarks.