Compare every major embedding model on retrieval quality, dimensions, context length, and API cost. Sortable. Filterable. No signup.
| Model ↕ | MTEB Retrieval ↓ | Dims ↕ | Max Tokens ↕ | $/1M tokens ↕ | Multi- lingual |
Open Source |
Best For |
|---|
Answer 4 questions and get a curated recommendation.
MTEB (Massive Text Embedding Benchmark) is the industry-standard leaderboard for embedding models. It tests retrieval, clustering, classification, and semantic similarity across dozens of datasets. The Retrieval score here is the most relevant for RAG and search — it measures nDCG@10 on real-world document retrieval tasks. Higher = better. Top models score 60–70. Scores above 65 are excellent.
Dimensions are the length of the embedding vector your model outputs (e.g., 1536 for OpenAI text-embedding-3-large). More dimensions generally capture more nuance — but they also increase storage cost, memory, and query latency in your vector database. Many modern models support Matryoshka representation, meaning you can truncate to fewer dimensions with minimal quality loss. Match dimensions to your retrieval accuracy vs. speed tradeoff.
Almost every production embedding pipeline uses cosine similarity (or its equivalent, dot product on normalized vectors) to compare embeddings. It measures the angle between vectors — not raw magnitude — which makes it robust to document length differences. You rarely need Euclidean distance or learned similarity functions. Normalize your vectors once at index time, then use dot product for maximum speed in any vector DB.