Google Launches First Native Multi-Modal Embedding Model Gemini Embedding 2, Supporting Multi-Type Data Processing

Google recently released the new Gemini Embedding 2 model, its first native multi-modal embedding model. Unlike generative models, this model focuses on data understanding—it converts text, images, videos, audio and documents into machine-readable vector forms, supporting semantic intent recognition in 100 languages.

The model handles various data types: text with up to 8192 tokens context window, up to 6 images per request, videos up to 120 seconds, etc. Application scenarios include retrieval-augmented generation, semantic search, sentiment analysis, etc. For example, it can quickly find key evidence in litigation forensics, improving retrieval precision. It is now available for public preview via Gemini API and Vertex AI.