Curiosity · AI Model
Gemini Embedding 001
Gemini Embedding 001 (gemini-embedding-001) is Google's March 2025 text embedding model, built on the Gemini foundation. It produces 3,072-dimensional vectors (with Matryoshka truncation to 1536/768/256), supports 100+ languages, and topped the public MTEB multilingual leaderboard at launch — making it Google's default embedding choice for RAG and search.
Model specs
- Vendor
- Family
- Gemini Embedding
- Released
- 2025-03
- Context window
- 2,048 tokens
- Modalities
- text
- Input price
- $0.15/M tok
- Output price
- n/a
- Pricing as of
- 2026-04-20
Strengths
- Topped the public MTEB multilingual leaderboard at launch
- Matryoshka representation learning — truncate vectors without re-embedding
- 100+ language coverage makes it the default for multilingual RAG
- Unified model replaces the older text-embedding-004 / text-multilingual-embedding-002 line
Limitations
- Text-only — no native vision or multimodal embedding yet
- 2K-token input limit — long documents require chunking
- Vector dimensionality (3072) is high — costs more memory than 1024-dim alternatives
Use cases
- RAG pipelines — retrieval over enterprise knowledge bases
- Semantic search and clustering across 100+ languages
- Code search and code-similarity retrieval
- Recommendation and deduplication systems
Benchmarks
| Benchmark | Score | As of |
|---|---|---|
| MTEB Multilingual (mean) | ≈68.3 | 2025-03 |
| MTEB English (mean) | ≈73.6 | 2025-03 |
| Code search (CodeSearchNet) | ≈high SOTA | 2025-03 |
Frequently asked questions
What is Gemini Embedding 001?
Gemini Embedding 001 (gemini-embedding-001) is Google's March 2025 text embedding model. It produces up to 3,072-dimensional vectors, supports 100+ languages, and was state-of-the-art on MTEB multilingual at launch — replacing earlier text-embedding models on Vertex AI and the Gemini API.
What is Matryoshka representation learning?
Matryoshka is a training technique where the model is optimised so that any prefix of the vector (e.g., first 768 or 256 dimensions) is still a usable embedding. For Gemini Embedding 001, that means you can truncate 3072-dim vectors to save storage without re-embedding your corpus.
How much does Gemini Embedding 001 cost?
As of April 2026, Gemini Embedding 001 is priced at roughly USD 0.15 per million input tokens on the Gemini API. Pricing may differ slightly on Vertex AI depending on region and batch options.
How does Gemini Embedding 001 compare to OpenAI embeddings?
On public MTEB multilingual at launch, Gemini Embedding 001 led OpenAI's text-embedding-3-large by a few points. For English-only use cases both are strong; for multilingual RAG Gemini Embedding 001 is often the stronger default.
Sources
- Google — Gemini Embedding model — accessed 2026-04-20
- Google Cloud — Gemini pricing — accessed 2026-04-20