Curiosity · AI Model
Jina Embeddings v3
Jina Embeddings v3 is an open-source multilingual embedding model from Jina AI. It supports task-specific LoRA adapters (retrieval, classification, separation, text-matching) and Matryoshka truncation down to 32 dimensions, making it a flexible default for teams that need on-prem embeddings without vendor lock-in.
Model specs
- Vendor
- Jina AI
- Family
- Jina Embeddings v3
- Released
- 2024-09
- Context window
- 8,192 tokens
- Modalities
- text
- Input price
- $0.02/M tok
- Output price
- n/a
- Pricing as of
- 2026-04-20
Strengths
- Open weights under a permissive research/commercial split — self-hostable
- Task LoRA adapters specialise the same backbone for retrieval, classification, and clustering
- Matryoshka truncation down to 32 dims for tight memory budgets
- 8k-token context window — longer than Cohere Embed v3 (512)
Limitations
- Requires own inference infrastructure for self-hosting
- Commercial-use licence terms should be read carefully for enterprise apps
- Lags voyage-3 and text-embedding-3-large on some English-only retrieval benchmarks
Use cases
- Self-hosted RAG on private clouds or on-prem GPUs
- Multilingual semantic search without API dependency
- Classification and clustering via task-specific LoRAs
- Edge deployments using Matryoshka-shrunk vectors
Benchmarks
| Benchmark | Score | As of |
|---|---|---|
| MTEB (multilingual avg) | ≈65.5 | 2024-09 |
| MIRACL (18 lang) | ≈58 | 2024-09 |
Frequently asked questions
What is Jina Embeddings v3?
Jina Embeddings v3 is an open-weight multilingual text embedding model from Jina AI. It supports 89 languages, an 8192-token context, Matryoshka truncation, and task-specific LoRA adapters.
Can I self-host Jina Embeddings v3?
Yes — the weights are published on Hugging Face and can be run locally with the transformers library or Jina's own inference stack. Jina also offers a hosted API if you prefer a managed endpoint.
What are the task LoRAs in Jina v3?
Jina v3 ships LoRA adapters for retrieval, text-matching, classification, and separation tasks. You select the right adapter per request, which changes the output embedding to suit that downstream use without retraining the base model.
How does Jina v3 compare to BGE-M3?
Both are open-weight multilingual embedding models with 8k context. Jina v3 uses task LoRAs and Matryoshka output, while BGE-M3 emphasises dense/sparse/multi-vector output in one model. Both are strong self-hosted defaults; pick based on retrieval style.
Sources
- Jina AI — Embeddings v3 announcement — accessed 2026-04-20
- Hugging Face — jinaai/jina-embeddings-v3 — accessed 2026-04-20