jina-embedding-b-en-v1

jina-embedding-b-en-v1 was Jina AI's first publicly released text embedding model: a 110M-parameter bidirectional encoder that maps English text into 768-dimensional vectors. Designed for semantic search, similarity comparison, and content recommendation at a time when 512-token context was the industry standard, it established Jina's entry into the open-source embedding space. The model is now legacy and has been superseded by the v2 and v3 families, which support 8K+ token contexts and multilingual input.

The model uses a T5-encoder architecture with mean pooling to produce fixed-length 768-dimensional representations. Training followed a two-phase contrastive recipe on the Linnaeus-Clean dataset (385M sentence pairs filtered from 1.6B candidates): first, InfoNCE loss on positive text pairs to learn semantic alignment; second, triplet loss to sharpen discrimination between similar but semantically distinct texts. A key design decision was the mean-pooling strategy, which averages token-level representations to produce a single vector capturing global sentence meaning. The 512-token context window was standard for its era but limits the model's usefulness for long-document applications.

On STS12, the model achieved a correlation score of 0.751, outperforming all-mpnet-base-v2 and all-minilm-l6-v2 at release. It demonstrated strong performance across standard English sentence-embedding tasks while maintaining fast inference times suitable for production. Its 512-token context window and English-only focus made it unsuitable for the long-document and multilingual workloads that became standard by 2025. The model has been superseded by jina-embeddings-v2-base-en (8K context, bidirectional ALiBi) and jina-embeddings-v3 (570M params, 89 languages, task-specific LoRA adapters).

This model is legacy and should not be used for new projects. If migrating from jina-embedding-b-en-v1, upgrade to jina-embeddings-v5-text-small for current workloads — it supports 32K token context, 89 languages, and task-specific LoRA adapters. For code-specific embedding needs, use jina-code-embeddings-1.5b. The model requires CUDA-capable hardware for optimal performance and accepts inputs up to 512 tokens. It is not suitable for multilingual content, long documents, or code-centric applications.