← All models

Gemini Embedding Experimental

Published Publisher Googlegemini-embedding-exp-03-07

Model positioning

An experimental text embedding model released by Google on March 7, 2025. It converts text into vectors to represent semantics rather than generating conversational answers; at release, it was available through the Gemini API's embed_content interface. (developers.googleblog.com)

Capabilities and uses

  • Supports more than 100 languages and can process content such as text and code, making it suitable for retrieval-augmented generation (RAG), semantic retrieval, recommendations, and text classification.
  • The release announcement stated an input limit of 8K tokens and an output vector of approximately 3K dimensions; it supports Matryoshka Representation Learning, allowing vectors to be truncated to reduce storage costs.
  • Google reported that at release, the model had an average task score of 68.32 on the MTEB multilingual leaderboard; this was a test result at release and does not represent its current ranking. (developers.googleblog.com)

Limitations and status

At release, the model was experimental, had limited capacity, and was subject to change. Google subsequently announced that the Gemini API would no longer support this experimental version as of August 14, 2025, and recommended migrating to gemini-embedding-001; the announcement stated that existing content would not need to be re-embedded when migrating. The specifications of the two versions should not be conflated. (developers.googleblog.com)

Pricing

As of November 19, 2026, no official input or output billing prices, currencies, or billing units applicable to this experimental version had been verified. The prices for gemini-embedding-001 listed on Google's current pricing page apply to a different model and cannot be directly applied to this experimental version. (ai.google.dev)

Official sources