Jina Embeddings v3 (Multilingual)

O Jina AI

Založeno 2020 · Berlin, Germany

Jina AI was founded in February 2020 in Berlin by Han Xiao (CEO, ex-Tencent and Zalando) together with co-founders Maximilian Werk, Christina Reher and Vincent Zhang. The company started as an open-source neural search framework (Jina, DocArray) and later pivoted to building proprietary multimodal embedding and reranking models offered through a hosted API. Jina has raised over $40M from investors including Canaan Partners, Mango Capital, Yunqi Partners and SAP, and ships its products under a freemium model. The Jina Embeddings v3 family was released in September 2024 and was the first open-weights embedding model to feature task-specific Low-Rank Adaptation (LoRA) heads selected at inference time, plus an 8,192-token context window. The release was widely covered as a top-tier multilingual embedding alternative on the MTEB leaderboard.

Architektura

Transformer bi-encoder with task-specific LoRA heads and Matryoshka representation learning

Jina Embeddings v3 is a 570M-parameter Transformer bi-encoder based on the XLM-RoBERTa architecture with several upgrades: rotary position embeddings, FlashAttention 2 and an extended context window of 8,192 tokens. It supports 89 languages with strong cross-lingual retrieval. A distinguishing feature is a set of five task-specific LoRA adapters (retrieval.query, retrieval.passage, separation, classification, text-matching) that are swapped at inference time by passing a 'task' parameter, which improves quality on each downstream task without retraining the base model. The output is a 1024-dimensional vector with Matryoshka representation learning, so truncation to 256 / 512 / 768 dimensions remains semantically meaningful and allows a quality vs. storage trade-off. Training used a multi-stage curriculum on multilingual text-pair data, search-query/document pairs and curated NLI data. Weights are released on Hugging Face under CC-BY-NC 4.0 for research use; commercial use is permitted via the hosted API or a paid commercial licence.

Parametry

570M

Kontext

8 192 tokenů

Schopnosti

• 89 languages with strong cross-lingual retrieval
• 8,192-token context window for long-document embedding
• Task-specific LoRA adapters (query, passage, classification, clustering, similarity)
• Matryoshka representation learning: truncate to 256/512/768 dims with graceful degradation
• Top-tier MTEB performance for a model under 1B parameters
• Open weights on Hugging Face for research; commercial via API or paid licence
• Best for: multilingual long-document RAG, semantic search, embedding-heavy SaaS

Trénování a licence

Multi-stage training on multilingual text pairs, search-query/document pairs, NLI data and curated synthetic data. Exact token count not disclosed.

Licence: Weights under CC-BY-NC 4.0 on Hugging Face (research only). Commercial use via Jina API or a paid commercial licence.

Bezpečnostní testy: No formal red-team report. Embeddings are not subject to content filters.