jina-embeddings-v2-base-zh

jina-embeddings-v2-base-zh is a 161M-parameter bilingual text embedding model for Chinese and English with an 8,192-token context window and 768-dimensional output. It was the first open-source model to seamlessly handle both Chinese and English with long-context support, addressing the unique tokenization challenges of Chinese character-based text in transformer architectures.

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-phase approach: initial pretraining on high-quality Chinese-English bilingual data, followed by primary and secondary fine-tuning stages with contrastive loss and hard-negative mining. The ALiBi mechanism enables the 8,192-token context window without learned positional embeddings. A notable improvement in the final release was a refined similarity score distribution that addressed score inflation issues present in the preview version, producing more discriminative and well-calibrated similarity scores.

On the C-MTEB (Chinese MTEB) leaderboard, the model demonstrated exceptional performance among models under 0.5GB, particularly excelling in Chinese-language tasks. It significantly outperformed OpenAI's text-embedding-ada-002 on Chinese-specific retrieval and similarity tasks while maintaining competitive performance on English tasks. The refined similarity score distribution improved discrimination between related and unrelated content in both languages. In 2026, jina-embeddings-v5-text-small supersedes this model with 89 languages, 32K context, and task-specific LoRA adapters.

Optimal for Chinese-English bilingual retrieval, cross-lingual document search, and multilingual content 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 multilingual projects, prefer jina-embeddings-v5-text-small` (89 languages, 32K context, LoRA adapters). CUDA-capable GPU recommended for production throughput. Input text should be in Chinese or English; the model handles both languages natively without translation.