Curiosity · AI Model

Jina Embeddings v4

Jina Embeddings v4 is the fourth-generation embedding model from Jina AI, released mid-2025. It is a 3.8-billion-parameter multimodal embedder built on a Qwen 2.5-VL backbone that handles text, images, and mixed visual documents (PDFs, screenshots) in 100+ languages. v4 supports Matryoshka Representation Learning — letting a client truncate a 2048-dim vector down to 128 dims at query time — and exposes a late-interaction ColBERT-style output alongside traditional single-vector retrieval. Released open under a Jina licence for research use.

Model specs

Vendor
Jina AI
Family
Jina Embeddings
Released
2025-06
Context window
32,768 tokens
Modalities
text, vision
Input price
$0.02/M tok
Output price
n/a
Pricing as of
2026-04-20

Strengths

  • Handles text, images, and visual documents in one model
  • Matryoshka dimensions — truncate to match storage budget
  • Single-vector and late-interaction modes
  • 100+ language coverage

Limitations

  • 3.8B params heavier than smaller Jina v3 / BGE embedders
  • Jina licence has some commercial restrictions
  • Higher VRAM than CPU-only embedders
  • Late-interaction mode needs retrieval engine support (e.g. Vespa / Qdrant)

Use cases

  • Visual-document RAG (PDFs, screenshots, charts)
  • Multilingual semantic search
  • Adjustable-dimension vector stores via Matryoshka
  • Hybrid late-interaction + dense retrieval

Benchmarks

BenchmarkScoreAs of
MTEB-Multilingual (avg)top-tier at release2025-06
ViDoRe visual-document retrievalbest-in-class single model2025-06

Frequently asked questions

What is Jina Embeddings v4?

Jina AI's fourth-generation multilingual multimodal embedding model, released mid-2025 with Matryoshka support and optional late-interaction output.

What is Matryoshka Representation Learning?

A training technique that makes the first N dimensions of an embedding independently useful, letting clients truncate long vectors to shorter ones without re-embedding.

Is Jina v4 open-source?

Weights are available on Hugging Face; commercial use depends on plan and licence tier with Jina AI.

Sources

  1. Jina Embeddings v4 announcement — accessed 2026-04-20
  2. Jina Embeddings v4 on Hugging Face — accessed 2026-04-20