jina-embeddings-v5-text-nano

jina-embeddings-v5-text-nano is a 239M-parameter multilingual text embedding model built on the EuroBERT-210M backbone. It supports 8K token context, produces 768-dimensional embeddings with Matryoshka truncation down to 32 dimensions, and delivers the highest accuracy-per-parameter in Jina's sub-500M embedding family. Designed for edge deployments, latency-sensitive applications, and resource-constrained environments where a full GPU is unavailable.

The model is built on EuroBERT-210M, a compact bidirectional encoder pretrained on multilingual data covering 15 major languages. It employs Last-Token-Pooling to generate embeddings from the final token representation. Training follows a two-stage distillation recipe: first, knowledge distillation from a larger teacher model transfers embedding quality; second, task-specific contrastive loss fine-tunes the model on retrieval, text-matching, clustering, and classification tasks. Matryoshka Representation Learning allows embedding dimensions to be truncated at any supported size (32, 64, 128, 256, 512, 768) while maintaining strong performance. Geometric Orthogonal Regularization (GOR) limits performance degradation under binary quantization to under 2 points on MTEB retrieval. The 8K context window enables long-document processing without chunking.

On MMTEB (multilingual), the model achieves 65.5 average (task-level) and 57.7 average (type-level) at just 239M parameters, outperforming all models under 500M including KaLM-mini-v2.5 (60.1, 494M), voyage-4-nano (58.9, 480M), and Gemma-300M (61.1, 308M). Task-level scores: classification 69.2, clustering 52.7, pair classification 81.9, reranking 64.6, retrieval 63.3, STS 78.2. On English MTEB, it achieves 71.0 average, nearly matching the much larger jina-embeddings-v5-text-small (71.7). Retrieval-specific: 63.26 on MTEB-M, 64.08 on RTEB, 56.06 on BEIR, 63.65 on LongEmbed. Embeddings remain robust under binary quantization thanks to GOR regularization.

Select the appropriate task prefix: 'retrieval' for asymmetric query-document search, 'text-matching' for symmetric similarity, 'clustering' for grouping, 'classification' for categorization. Matryoshka truncation to 256 dimensions preserves over 95% of retrieval quality while reducing storage by 67%. Binary quantization is supported for further storage reduction. The EuroBERT backbone provides strong coverage for 15 major languages including English, French, German, Spanish, Chinese, Japanese, Arabic, and Hindi. Use cosine similarity for embedding comparison. Available via Jina AI API, Hugging Face (Sentence Transformers, vLLM), and quantized variants for llama.cpp. For workloads requiring 32K context or higher accuracy, use jina-embeddings-v5-text-small instead.