jina-embeddings-v3

jina-embeddings-v3 is a 570M-parameter multilingual text embedding model supporting 89 languages and an 8,192-token context window. Its defining innovation is the use of task-specific LoRA adapters: the same 570M backbone produces optimized embeddings for retrieval, classification, clustering, and text-matching by swapping lightweight adapter weights (each under 60M parameters). It achieved state-of-the-art MTEB performance at release, outperforming OpenAI and Cohere proprietary embeddings on English tasks.

The model is built on a jina-XLM-RoBERTa backbone with 24 transformer layers, using Rotary Position Embeddings (RoPE) with a tuned base frequency for long-context extrapolation up to 8,192 tokens. Training followed a three-stage curriculum: (1) multilingual pretraining on 89 languages, (2) contrastive fine-tuning on paired-text datasets to establish base embedding quality, and (3) task-specific LoRA adapter training. Each LoRA adapter adds less than 3% to the total parameter count while specializing the model for a specific task: retrieval.query and retrieval.passage for asymmetric search, separation for clustering, classification for categorization, and text-matching for semantic similarity. Matryoshka Representation Learning (MRL) enables embedding truncation from 1024 to 32 dimensions while preserving over 90% of retrieval quality at 64 dimensions. Late chunking is supported for documents exceeding the context window.

On MTEB, jina-embeddings-v3 achieved an average score of 65.52 across all tasks, with particularly strong results in classification (82.58) and sentence similarity (85.80). It outperformed multilingual-e5-large-instruct across all multilingual tasks and beat e5-mistral-7b-instruct (12× the parameter count) on English tasks. On multilingual MTEB, it scored 64.44, demonstrating consistent performance across 89 languages. MRL truncation preserved 92% of retrieval performance at 64 dimensions versus the full 1024. In 2026, jina-embeddings-v5-text-small supersedes this model for most workloads, offering 32K context and distillation-based training that yields higher accuracy per parameter.

Select the appropriate LoRA adapter for your task: retrieval.query/retrieval.passage for search, separation for clustering, classification for categorization, text-matching for semantic similarity. Use the `late_chunking parameter for documents exceeding 8,192 tokens. Matryoshka truncation to 256 or 512 dimensions is suitable for storage-constrained indexing; use full 1024 dimensions for re-ranking. The model requires CUDA-capable GPU for production throughput. For new multilingual projects, prefer jina-embeddings-v5-text-small` (32K context, distillation-trained, higher accuracy). The model is not designed for text generation or complex reasoning — it produces embeddings, not text.