Gemini Embedding 2
2026-03-10 · Publisher : Google · Domain : Artificial Intelligence
Google DeepMind launches Gemini Embedding 2, the first natively multimodal embedding model, which places text, images, video, audio and documents in a single vector space, in public preview.
What changes
- A single embedding space for text, images, videos, audio and PDFs, with interleaved inputs.
- More than 100 languages supported.
- Matryoshka representation: default dimension 3072, reducible to 1536 or 768.
- Simplifies multimodal RAG, semantic search and clustering pipelines.
### Specifications - Input limits: 8192 tokens of text, 6 images, 120 seconds of video, 6 PDF pages per request. - Availability: Gemini API and Vertex AI, first in public preview, then moved to general availability (GA date not verified). - Identifier: gemini-embedding-2-preview at announcement. ### Pricing | Modality | Standard (USD / 1M) | Batch (USD / 1M) | |---|---|---| | Text | 0.20 | 0.10 | | Image | 0.45 | 0.225 | | Audio | 6.50 | 3.25 | | Video | 12.00 | 6.00 | Pricing recorded from the current pricing page. ### Benchmarks No score table included here; the post indicates a new state of the art in text, image, video and speech without detailing the figures in the extracted text. ### Key takeaways A multimodal replacement for the previous text embedding model, useful for a RAG covering documents, images and media with a single index.
Links
Tags : LLM, Google, Gemini, Embedding, RAG