Gemini Embedding

2025-03-07 · Publisher : Google · Domain : Artificial Intelligence

Google launches Gemini Embedding, a text embedding model derived from Gemini: an experimental version on March 7, 2025, ranked first on the multilingual MTEB leaderboard, then a stable version on July 14.

Google launches Gemini Embedding, a text embedding model derived from Gemini: an experimental version on March 7, 2025, ranked first on the multilingual MTEB leaderboard, then a stable version on July 14. ### What changes - Embedding model trained from Gemini, inheriting its understanding of language and context. - First place on the MTEB (Multilingual) leaderboard with a mean score of 68.32, 5.81 points ahead of the next model, according to Google. - Stable version `gemini-embedding-001` on July 14, 2025, available in the Gemini API and Vertex AI. - More than 100 languages, outputs that can be reduced through Matryoshka representation learning (3,072 dimensions by default). ### Specifications - Maximum input: 2,048 tokens (stable version). - Output dimensions: 3,072 by default, reducible (3,072, 1,536 or 768 recommended). - Identifiers: `gemini-embedding-exp-03-07` (experimental, discontinued in 2025) then `gemini-embedding-001`. ### Pricing | Model | Input (USD / 1M tokens) | |---|---| | gemini-embedding-001 | 0.15 | Price announced at general availability on July 14, 2025. Free and paid tiers in the Gemini API. ### Benchmarks | Benchmark | Score | |---|---| | MTEB Multilingual (task mean, experimental version) | 68.32 | Scores published by the vendor. ### Key takeaways A leading multilingual embedding model for semantic search and RAG, with adjustable dimensions to control storage cost.

Tags : LLM, Google, Gemini, Embeddings, RAG