E5-base-v2

E5-base-v2 is Microsoft's workhorse English text embedding model, part of the E5 family from the unilm project. With 12 layers, 109 million parameters and 768-dimensional embeddings, it delivers strong retrieval performance on the BEIR benchmark while remaining small enough to run anywhere — a go-to default for semantic search, clustering and classification pipelines.\n\nThe v2 generation improved on the original E5 models with better contrastive pretraining, and the model is a staple baseline in embedding benchmarks and RAG evaluations. It is distributed under the MIT license, making it free for commercial use.\n\nUsage follows a simple convention: prefix inputs with 'query: ' or 'passage: ' for retrieval tasks, or 'query: ' alone for other tasks, then average-pool and normalize the outputs.