E5 Large v2

E5 Large v2 is a text embedding model that converts English sentences, paragraphs, and documents into high-dimensional numerical representations. Developed by Microsoft and maintained by researcher Liang Wang under the intfloat organization, the model uses a 24-layer transformer architecture that produces 1024-dimensional embeddings. Unlike general-purpose language models, this embedding specialist relies on a unique input format requiring "query: " or "passage: " prefixes to distinguish search queries from searchable content, which helps the model understand its role in retrieval tasks. This design makes it particularly effective for semantic search, retrieval augmented generation pipelines, and document clustering operations.

The model builds on weakly-supervised contrastive pre-training, learning from large-scale unlabeled data pairs rather than relying on manually curated labels. This training approach, detailed in the associated research paper, allows the model to handle messy, real-world data and short queries that return medium-length passages. Benchmarks on BEIR and MTEB show strong performance gains over smaller variants like e5-base-v2, which uses only 12 layers and 768-dimensional embeddings. The model is distributed as open weights under the Apache 2.0 license, enabling developers to run, fine-tune, and deploy it without licensing restrictions across libraries like Hugging Face Transformers and Sentence Transformers.