jina-embeddings-v4

Name: jina-embeddings-v4
Author: jinaai

by jinaai

Universal multimodal multilingual embeddings with task-specific LoRA adapters

532Kdl/month

535likes

3.8Bparams

HuggingFace Run it on your own data, free

Identifiers

Model ID

jinaai/jina-embeddings-v4

Feature URI

mixpeek://image_extractor@v1/jina_embeddings_v4

Overview

Jina Embeddings v4 is a 3.8B-parameter multimodal embedding model built on the Qwen2.5-VL-3B-Instruct backbone. It unifies text and image representations through a shared pathway, supporting both single-vector (2048-dim, truncatable to 128) and multi-vector (128-dim per token) output modes for late-interaction retrieval.

Three task-specific LoRA adapters (60M parameters each) optimize performance for retrieval, text-matching, and code search without modifying the frozen backbone. On Mixpeek, jina-embeddings-v4 powers cross-modal search across documents with tables, charts, and mixed-media content, excelling where visual layout matters as much as text.

Architecture

Qwen2.5-VL-3B-Instruct backbone with vision encoder for image-to-token conversion. Dual output modes: single-vector (2048-dim via mean pooling) and multi-vector (128-dim per token via projection layers). Three frozen LoRA adapters (60M each) for retrieval, text-matching, and code search tasks.

Mixpeek SDK Integration

// No extractor parameter takes a Hugging Face model id (checked against
// GET /v1/discovery/extractors, which returns 13), so jina-embeddings-v4 runs
// on your side and the output is upserted through POST
// /v1/namespaces/{namespace_id}/documents/upsert. On Enterprise the other
// path is to upload the weights instead: POST /v1/namespaces/{id}/models
// accepts the huggingface format and a custom plugin loads them.
const res = await fetch(
  "https://api.mixpeek.com/v1/namespaces/ns_your_namespace/documents/upsert",
  {
    method: "POST",
    headers: {
      Authorization: "Bearer API_KEY",
      "Content-Type": "application/json",
    },
    body: JSON.stringify({
      collection_id: "col_your_collection",
      documents: [
        {
          document_id: "asset-00412",
          // The vector name has to match a vector index on the collection.
          vectors: { "image-embedding": yourVector },
          payload: { source_key: "archive/2026/asset-00412" },
        },
      ],
    }),
  },
);

// Managed alternative, if this exact model is not the requirement:
// image_extractor@v1 runs google/siglip-base-patch16-224
// (768-d) over a bucket, with no inference of your own.

Capabilities

Multimodal: text and image in a shared embedding space
2048-dimensional single-vector or 128-dim multi-vector output
Task-specific LoRA adapters for retrieval, matching, and code
Matryoshka dimensions (2048 down to 128)
Strong on visually rich documents: tables, charts, diagrams

Use Cases on Mixpeek

Cross-modal document retrieval where layout and visuals matter (charts, infographics)

Multilingual semantic search across mixed-media collections

Code search and retrieval with the dedicated code LoRA adapter

Benchmarks

Dataset	Metric	Score	Source
MTEB-en (text retrieval)	nDCG@10	55.97	Jina AI, 2025: jina-embeddings-v4 paper
CLIP Benchmark (cross-modal)	Score	84.11	Jina AI, 2025: jina-embeddings-v4 paper
LongEmbed	Score	67.11	Jina AI, 2025: jina-embeddings-v4 paper