jina-embeddings-v2-base-en

jina-embeddings-v2-base-en is a 137M-parameter English text embedding model with an 8,192-token context window — 16× the standard 512-token limit of its era. Built on a BERT-small backbone with symmetric bidirectional ALiBi (Attention with Linear Biases), it was the first open-source Jina embedding to natively handle long documents without truncation or chunking. It produces 768-dimensional vectors and remains a solid choice for English-only retrieval where multilingual or long-context v3/v5 features are not needed.

The architecture pairs a BERT-small transformer (12 layers, 12 attention heads, 768 hidden dimensions) with symmetric bidirectional ALiBi positional encodings. ALiBi replaces learned positional embeddings with a linearly decaying attention bias, allowing the model to extrapolate far beyond its 512-token training length up to 8,192 tokens without performance degradation. Training followed a two-stage pipeline: pretraining on C4, then fine-tuning on Jina's curated collection of 40+ specialized sentence-pair datasets with hard-negative mining. Symmetric bidirectional attention ensures each token attends to both preceding and following context, producing representations that capture global sentence meaning. Mean pooling over all token representations produces the final 768-dimensional embedding.

At release, the model outperformed OpenAI's text-embedding-ada-002 on multiple MTEB English subtasks: classification (73.45% vs. 70.93%), reranking (85.38% vs. 84.89%), retrieval (56.98% vs. 56.32%), and summarization (31.6% vs. 30.8%). Its 8,192-token context was a significant advantage over competing models limited to 512–2,048 tokens, enabling document-level retrieval without chunking. The compact 307MB footprint made it deployable on consumer-grade GPUs. In 2026, jina-embeddings-v5-text-small (677M params, 32K context, task-specific LoRA adapters) outperforms it for most production workloads, but it remains relevant for lightweight English-only pipelines.

Use this model for English-only retrieval where the 8K context window is sufficient and multilingual or task-specific adapters are not required. For documents exceeding 8,192 tokens, apply semantic chunking before embedding. The model integrates with major vector databases (Qdrant, Weaviate, MongoDB Atlas, Milvus) and RAG frameworks (LangChain, LlamaIndex, Haystack). For new projects requiring multilingual support, 32K context, or task-specific optimization, prefer jina-embeddings-v5-text-small. CUDA-capable GPU recommended for production throughput; CPU inference is possible but significantly slower.