Jina AI models

State-of-the-art models for each stage of the retrieval pipeline

Purpose-built for retrieval, Jina AI models deliver accuracy and speed that outperforms models 5× their size. Multilingual, multimodal, and now native on Elasticsearch.

Try Jina on Elasticsearch
Learn more

Meet the Jina AI models

Our frontier models form the search foundation for high-quality enterprise search and retrieval augmented generation (RAG) systems.

  • Reader

    Convert complex documents, web pages, and PDFs into clean, structured input for search and large language models (LLMs).

  • Embeddings

    Improve search and RAG systems with multimodal and multilingual embeddings for text, images, and code.

  • Reranker

    Maximize relevance with a world-class reranker that delivers precision for critical applications like RAG, AI assistant, and agents.

Compact by design, precise by results

Go from raw data to high-precision results in one API.

  • Multimodal search, 30 languages

    Jina's models work across text, images, and code. Jina-vlm makes visual content searchable alongside text and documents. 30+ languages are supported natively, and cross-language search works out of the box.

  • Best results, not just nearest

    Jina's reranking models are proven leaders. Get extra precision with rerankers that rescore every candidate against the original query, using deep analysis to get the most relevant answers on top.

  • Smart training, smaller models

    Jina's models are trained on tasks that matter for retrieval: finding the right document and best answer from messy sources. That's why they match or outperform larger models at a fraction of the cost.

  • Map any field as semantic_text and Elasticsearch generates embeddings automatically. On EIS, Jina models default to deliver out-of-the-box multilingual and multimodal semantic search with zero config.

  • One API call, that's all

    Combine traditional keyword search with Jina's semantic matching in a single query. Use one API call with reciprocal rank fusion to merge the best of each approach.

  • Lean at any scale

    Combine Jina variable-sized embeddings with Elastic's vector quantization (BBQ) to reduce storage by up to 95% with minimal accuracy loss. Turn precision all the way up when accuracy matters the most.

Our research

Jina's models are built on research presented at top machine learning (ML) conferences, including CVPR, NeurIPS, and EMNLP. Explore how our frontier search models were trained from scratch in our latest publications.

  • Jina-embeddings-v5-text: Task-Targeted Embedding Distillation

    We introduce a novel training regimen that combines model distillation techniques with task-specific contrastive loss to produce compact, high-performance embedding models.

  • Embedding Inversion via Conditional Masked Diffusion Language Models

    We frame embedding inversion as conditional masked diffusion, recovering all tokens in parallel through iterative denoising rather than sequential autoregressive generation.

  • Embedding Compression via Spherical Coordinates

    We present a compression method for unit-norm embeddings that achieves 1.5× compression, 25% better than the best prior lossless method.

Join our open source community

Jina's models are open-weight and freely available on Hugging Face, with millions of monthly downloads. The codebase is public on GitHub. The community has direct access to our developers.

Frequently asked questions

What are Jina search models?

Jina models are open source, frontier AI models for retrieval. They include embedding models for vectors, rerankers for precision, and readers for extracting and structuring content from URLs and docs.

Try it now on Elastic Cloud

Spin up a fully loaded deployment on the cloud provider you choose. As the company behind Elasticsearch, we bring our features and support to your Elastic clusters in the cloud.

Start free trial