Gemini Embedding 2: First Natively Multimodal Embedding Model.

Google launched Gemini Embedding 2 — text (8,192 tokens), images (6 per request), video (120s), and audio without transcription, producing a single 3,072-dimensional vector. Scores 68.32 on MTEB English (5.09-point margin over previous leaders). Natively supports Matryoshka truncation: cut to 768 dimensions with under 0.5% quality loss. For mixed-media RAG, this eliminates separate embedding pipelines per modality.