Gemini Embedding

weights

closed(no checkpoint distributed for gemini-embedding-2 or gemini-embedding-001

data

closed(training corpus not released

code

closed(no training or inference implementation

license

proprietary(hosted API under Google’s terms

Both gemini-embedding-2 and gemini-embedding-001 are served only through the Gemini API and Vertex AI, with no downloadable weights, training data or code. The published technical report makes the benchmark numbers auditable, but a paper is not a released corpus, checkpoint or pipeline, so the score reflects what actually ships rather than what is described.

• https://ai.google.dev/gemini-api/docs/embeddings recorded 2026-09-11

  'The latest model, gemini-embedding-2, is the first multimodal' embedding model, and 'for text-only use cases, gemini-embedding-001 remains available'. Both are invoked as embedContent calls against generativelanguage.googleapis.com; no download is offered.

• https://ai.google.dev/gemini-api/docs/pricing recorded 2026-09-11

  The Gemini API pricing page lists Gemini Embedding 2 (gemini-embedding-2) and Gemini Embedding (gemini-embedding-001) as priced hosted models, which is the only way either is sold.

• https://arxiv.org/abs/2605.27295 recorded 2026-09-11

  The Gemini Embedding 2 report describes 'large-scale contrastive learning in a multi-task multi-stage training setup' and releases no corpus, checkpoint or training code.

Google publishes no call-volume figure for Gemini Embedding, so adoption is read from distribution: general availability on the Gemini API and Vertex AI, a free tier alongside paid limits, and Google's migration of users off the deprecated embedding-001 and text-embedding-004 models onto this line.

• https://ai.google.dev/gemini-api/docs/pricing recorded 2026-09-11

  Both gemini-embedding-2 and gemini-embedding-001 carry list prices on the Gemini API pricing page, so both are generally sold rather than previewed.

• https://developers.googleblog.com/en/gemini-embedding-available-gemini-api/ recorded 2026-09-11

  gemini-embedding-001 'is now generally available to developers in the Gemini API and Vertex AI'; free and paid tiers are offered, and the legacy embedding-001 and text-embedding-004 models were deprecated on 2025-08-14 and 2026-01-14 respectively, migrating their users onto this line.

Gemini Embedding 2 reports real, published numbers across several benchmarks, and its multimodal span across text, image, video and audio is wider than any peer's. It falls short of the frontier on two counts: 69.9 on MTEB multilingual trails Harrier-OSS's 74.3 on MMTEB v2, and Voyage's independent RTEB benchmark puts voyage-4-large 3.87% ahead of the older gemini-embedding-001.

• https://arxiv.org/abs/2605.27295 recorded 2026-09-11

  Abstract: 'a score of 62.9 R@1 on MSCOCO, 68.8 NDCG@10 on Vatex, 69.9 on MTEB multilingual and 84.0 on MTEB Code ... surpassing the performance of specialized models'.

• https://blog.voyageai.com/2026/01/15/voyage-4/ recorded 2026-09-11

  Over all 29 RTEB datasets, voyage-4-large surpasses Gemini Embedding 001 by an average of 3.87%.

• https://developers.googleblog.com/en/gemini-embedding-available-gemini-api/ recorded 2026-09-11

  gemini-embedding-001 'has consistently held a top spot on the Massive Text Embedding Benchmark (MTEB) Multilingual leaderboard'; supports over 100 languages with a 2,048-token maximum input and 3,072 default output dimensions.