jina-embeddings-v2-base-es

jina-embeddings-v2-base-es is a 161M-parameter bilingual text embedding model for Spanish and English, featuring an 8,192-token context window and 768-dimensional output. It was designed for cross-lingual retrieval in Spanish-speaking markets, mapping semantically equivalent Spanish and English content into a shared embedding space. The model addresses a critical gap: most embedding models at the time were English-centric, with Spanish performance significantly degraded.

The model uses a BERT-based backbone with symmetric bidirectional ALiBi positional encodings, 161M parameters, and a 768-dimensional output space. Training followed a three-stage process: (1) pretraining on Spanish-English parallel corpora, (2) contrastive fine-tuning with hard-negative mining on curated sentence pairs, and (3) cross-lingual alignment to ensure Spanish and English representations of the same concept cluster together. The ALiBi mechanism enables the 8,192-token context without learned positional embeddings, allowing the model to extrapolate beyond its 512-token training length. Mean pooling produces the final embedding vector.

The model outperformed significantly larger multilingual models (E5, BGE-M3) in Spanish-English retrieval tasks while being only 15–30% of their size. It demonstrated strong performance on MTEB Spanish subtasks, particularly in retrieval and clustering. The 8,192-token context window provided consistent performance on multi-page documents where most competing models required chunking. In 2026, jina-embeddings-v5-text-small supersedes this model, offering 89 languages, 32K context, and task-specific LoRA adapters. The v2-base-es model remains a cost-effective option for dedicated Spanish-English pipelines.

Ideal for Spanish-English bilingual search, content recommendation, and cross-lingual document analysis. For documents exceeding 8,192 tokens, use semantic chunking or the `late_chunking parameter via the Jina API. The model integrates with major vector databases and RAG frameworks. For new projects requiring multilingual coverage beyond Spanish-English, 32K context, or task-specific optimization, prefer jina-embeddings-v5-text-small`. CUDA-capable GPU recommended for production. Input text should be in Spanish or English; mixed-language input is supported but performance is optimized for the two trained languages.