NEWVectors or files. Pick a path.Start →
    Models/jinaai/jina-embeddings-v3
    CC-BY-NC-4.0

    jina-embeddings-v3

    by jinaai

    Jina AI's multilingual text embedding model: 1,024 dimensions you can truncate to 32, 8,192 tokens of context, and five task adapters, under a non-commercial licence

    Identifiers
    Model ID
    jinaai/jina-embeddings-v3

    Deploy jina-embeddings-v3

    Single-tenant

    Mixpeek has no managed extractor for this model. On a single-tenant deployment you upload the weights and a custom plugin serves them next to the rest of your pipeline.

    Overview

    jina-embeddings-v3 is a multilingual text embedding model from Jina AI, released in September 2024. It has about 572M parameters, reads up to 8,192 tokens, and returns a 1,024-dimension vector that can be cut to 768, 512, 256, 128, 64 or 32 dimensions (Matryoshka representation learning), so the stored width is chosen at index time. Five LoRA adapters tune the same base model for the task at hand: retrieval queries, retrieval passages, clustering, classification and text similarity.

    The weights are published under CC BY-NC 4.0, which allows research and evaluation and rules out commercial use without a licence from Jina AI. It has about 1.6 million downloads a month on Hugging Face.

    Architecture

    A 24-layer XLM-RoBERTa-style encoder (hidden size 1,024) with rotary position embeddings, which is what extends it to 8,192 tokens, and mean pooling over the token outputs. Five LoRA adapters sit on top of the shared weights, one per task: retrieval.query, retrieval.passage, separation, classification and text-matching. The task argument picks the adapter at encode time, so one checkpoint serves all five. Trained for Matryoshka embeddings at 32, 64, 128, 256, 512, 768 and 1,024 dimensions. The custom code loads with trust_remote_code=True.

    Mixpeek SDK Integration

    # Embed passages with jina-embeddings-v3, using the passage adapter.
    # pip install -U transformers torch einops
    from transformers import AutoModel
    
    model = AutoModel.from_pretrained("jinaai/jina-embeddings-v3", trust_remote_code=True)
    
    passages = ["Bucket objects keep the metadata you send.", "Retrievers chain search, filter and rank stages."]
    doc_vecs = model.encode(passages, task="retrieval.passage")   # (2, 1024), normalized
    # Store 256 dimensions instead of 1024 (Matryoshka):
    small = model.encode(passages, task="retrieval.passage", truncate_dim=256)

    Capabilities

    • Multilingual text embeddings (the model card lists 94 languages)
    • 8,192-token context per input
    • Matryoshka truncation from 1,024 down to 32 dimensions
    • Task adapters for query and passage retrieval, clustering, classification and similarity
    • Loads with Hugging Face Transformers and sentence-transformers

    Use Cases on Mixpeek

    Multilingual semantic search where queries and documents are in different languages
    Asymmetric retrieval for RAG, with queries and passages encoded by their own adapters
    Trading index size for accuracy by truncating stored vectors (512 or 256 dimensions)
    Research and evaluation baselines (the licence rules out commercial use without Jina's permission)

    Performance

    Input SizeUp to 8,192 tokens of text
    Embedding Dim1024 (truncatable to 768, 512, 256, 128, 64 or 32)
    GPU LatencyInput dependent
    GPU ThroughputBatch dependent
    GPU MemoryModel dependent

    We have not measured it. Queries and passages use their own adapters (retrieval.query and retrieval.passage), which are trained as a pair for asymmetric search. The model card records a fix to truncation, which now happens before normalization; vectors made with the older code revision are not comparable with new ones. At 1,024 dimensions in float32 a vector is 4,096 bytes, about 4.1 GB per million; truncated to 256 it is a quarter of that. CC BY-NC 4.0: commercial use needs a licence from Jina AI.

    Frequently Asked Questions

    What is jina-embeddings-v3?

    A multilingual text embedding model from Jina AI with about 572M parameters, released in September 2024. It reads up to 8,192 tokens and returns a 1,024-dimension vector that can be truncated to as few as 32 dimensions, and five LoRA adapters tune it for retrieval queries, retrieval passages, clustering, classification or text similarity.

    Can I use jina-embeddings-v3 commercially?

    Not under its published licence. The weights on Hugging Face are CC BY-NC 4.0, which allows non-commercial use with attribution, so a product needs a commercial licence from Jina AI or access through their paid API. For an open licence with similar reach, bge-m3 (MIT, 8,192 tokens, multilingual) and multilingual-e5-large-instruct (MIT) are the usual alternatives.

    What are the task adapters and which one should I use?

    Five LoRA adapters on one shared model, chosen with the task argument at encode time. For search, encode queries with retrieval.query and documents with retrieval.passage, since the two sides are trained to meet. Use text-matching for symmetric similarity between two texts of the same kind, separation for clustering and reranking, and classification for features fed to a classifier.

    How small can I make its vectors?

    Down to 32 dimensions, because it was trained with Matryoshka representation learning at 32, 64, 128, 256, 512, 768 and 1,024. Pass truncate_dim to encode, or cut the stored vector and renormalize. Each halving roughly halves storage and search cost, and quality falls gradually, so test 512 and 256 on your own queries before choosing. Truncate every vector in an index to the same length.

    How do I use it with Mixpeek?

    Mixpeek's managed text extractor runs multilingual-e5-large-instruct, not this checkpoint. To use jina-embeddings-v3, embed on your side under a licence that allows your use and upsert the vectors into a collection, or register it as a custom model on Enterprise. Keep queries on the same model, adapter and dimension as the stored vectors.

    Specification

    Organizationjinaai
    Retriever-
    Parameters572M
    LicenseCC-BY-NC-4.0
    Downloads/moN/A
    Likes1,174

    Research Paper

    Model paper or technical report

    arxiv.org

    Build a pipeline with jina-embeddings-v3

    Add this model to a processing pipeline alongside other extractors. Combine with retrieval stages for end-to-end search.

    Run it on your own data, free