Jina models

This page collects all Jina models you can use as part of the Elastic Stack.

Currently, the following models are available as built-in models:

Embedding models

• jina-embeddings-v5-omni-small — multimodal (text, image, audio, video, PDF)
• jina-embeddings-v5-omni-nano — multimodal (text, image, audio, video, PDF)
• jina-embeddings-v5-text-small — text-only
• jina-embeddings-v5-text-nano — text-only
• jina-embeddings-v3 — text-only

Rerankers

• jina-reranker-v3
• jina-reranker-v2

Embedding models convert text into vector embeddings, which are fixed-length numerical representations that capture semantic meaning. Texts with similar meaning are mapped to nearby points in vector space, so you can retrieve relevant documents with vector similarity search. When you send text to an EIS inference endpoint that uses an embedding model, the model returns a vector of floating-point numbers (for example, 1024 values). Elasticsearch stores these vectors in dense_vector fields or through the semantic_text field and uses vector similarity search to retrieve the most relevant documents for a given query. Unlike ELSER, which expands text into sparse token-weight vectors, these models produce compact dense vectors that are well suited for multilingual and cross-domain use cases.

The jina-embeddings-v5-omni-* models accept text, image, audio, video, and PDF inputs and place all supported input types in a shared vector space. Use them when you need cross-modal retrieval, such as querying a text index with an image or finding videos from a text query.

The v5 omni models are available through Elastic Inference Service (EIS), so no machine learning node scaling or model deployment is required.

jina-embeddings-v5-omni-small is the recommended Jina embedding model for deployments that need higher-quality mixed-media search. It produces 1024-dimension embeddings by default, supports a 32768 token input context window, and uses the same text embedding space as jina-embeddings-v5-text-small.

For more information about the model, refer to the Elastic blog post or the model page.

jina-embeddings-v5-omni-nano is the compact, lower-cost member of the Jina v5 omni family. It produces 768-dimension embeddings by default, supports a 32768 token input context window, and uses the same text embedding space as jina-embeddings-v5-text-nano.

For more information about the model, refer to the Elastic blog post or the model page.

To use a v5 omni model, you must have the appropriate subscription level. Elastic Cloud trial accounts cannot use the v5 omni models; start a paid Elastic Cloud deployment or Serverless project to access them.

All input types require Elastic Stack 9.5 or later.

For text input, the recommended entry point is a semantic_text field that references one of the preconfigured v5 omni inference endpoints. Elasticsearch provisions the endpoint on first reference.

Create an index with a semantic_text field:

				PUT multimodal-semantic-index
					{
  "mappings": {
    "properties": {
      "content": {
        "type": "semantic_text",
        "inference_id": ".jina-embeddings-v5-omni-small"
      }
    }
  }
}
		

Index documents normally. Elasticsearch generates embeddings through the inference endpoint:

				POST multimodal-semantic-index/_doc
					{
  "content": "'Kraft Dinner' is what Canadians call macaroni and cheese when prepared from a kit."
}
		

Query the field with a semantic query:

				GET multimodal-semantic-index/_search
					{
  "query": {
    "semantic": {
      "field": "content",
      "query": "Was bedeutet 'Kraft Dinner' für Kanadier?"
    }
  }
}
		

To use jina-embeddings-v5-omni-nano, set inference_id to .jina-embeddings-v5-omni-nano instead.

To create an explicit inference endpoint instead of using the preconfigured endpoint, use the embedding task type:

				PUT _inference/embedding/eis-jina-embeddings-v5-omni-small
					{
  "service": "elastic",
  "service_settings": {
    "model_id": "jina-embeddings-v5-omni-small"
  }
}
		

semantic_text ingests text content. To embed image, audio, video, or PDF input, or to issue a cross-modal query against a text index, call the inference endpoint directly and store or compare the resulting vector against a dense_vector field.

The request body is a structured input array. Each element holds a content object describing one piece of media, and each request can hold up to 16 input items. Media values are base64-encoded data URIs:

				POST _inference/embedding/.jina-embeddings-v5-omni-small
					{
  "input": [
    { "content": { "type": "image", "format": "base64", "value": "data:image/png;base64,iVBORw0KGgo..." } },
    { "content": { "type": "audio", "format": "base64", "value": "data:audio/wav;base64,UklGRiQAAAB..." } },
    { "content": { "type": "video", "format": "base64", "value": "data:video/mp4;base64,AAAAIGZ0eXA..." } },
    { "content": { "type": "pdf",   "format": "base64", "value": "data:application/pdf;base64,JVBE..." } }
  ]
}
		

To combine several media items into a single embedding, pass an array of content fields under one input element:

				POST _inference/embedding/.jina-embeddings-v5-omni-small
					{
  "input": [
    {
      "content": [
        { "type": "text",  "value": "A description of the scene" },
        { "type": "image", "format": "base64", "value": "data:image/png;base64,iVBORw0KGgo..." },
        { "type": "audio", "format": "base64", "value": "data:audio/wav;base64,UklGRiQAAAB..." }
      ]
    }
  ]
}
		

The response is shaped {"embeddings": [{"embedding": [...]}, ...]}. The array length matches the number of input items, except for PDF input, which produces one embedding per page.

The v5 omni models share their text embedding space with the matching v5 text models. Existing dense_vector data populated by jina-embeddings-v5-text-* remains directly comparable to vectors produced from text input by the corresponding omni model, so no reindex is required.

Use the matching pair:

• jina-embeddings-v5-omni-small with jina-embeddings-v5-text-small (1024 dimensions)
• jina-embeddings-v5-omni-nano with jina-embeddings-v5-text-nano (768 dimensions)

Do not mix across the small and nano families, because their vector spaces and dimensions differ.

For semantic_text mappings, set the inference_id to the corresponding .jina-embeddings-v5-omni-* endpoint on new indices. Existing indices continue to work unchanged.

For code that calls _inference directly, the task type changes from text_embedding to embedding, and the request body changes from a flat string input to an array of content objects. See Multimodal ingestion and querying.

For pure text workloads, keep using v5-text-* endpoints. Use v5-omni-* when mixed media is in scope.

• Use jina-embeddings-v5-omni-small when retrieval quality is the main priority. Use jina-embeddings-v5-omni-nano when ingestion volume, latency, or cost is the main constraint.
• Each inference request can contain up to 16 input items.
• Image inputs must be at least 28×28 pixels (784 pixels total).
• PDF inputs return one embedding per page.
• Video is sampled at 32 uniformly spaced frames regardless of clip length. For long videos, segment into shorter clips for finer temporal resolution.
• Although the models support a 32768 token context window, consider chunking very large text fields to control latency and cost.

The jina-embeddings-v5-text-small model is a compact, multilingual dense vector embedding model that you can use through Elastic Inference Service (EIS). It is optimized for retrieval, text matching, clustering, and classification with task-specific adapters and includes support for Matryoshka Representation Learning, which enables you to truncate embeddings to fewer dimensions with minimal loss in quality. As the model runs on EIS, Elastic's own infrastructure, no ML node scaling and configuration is required to use it.

The jina-embeddings-v5-text-small model has 677M parameters, supports a 32768 token input context window, and produces 1024-dimension embeddings by default.

For more information about the model family, refer to the Elastic blog post or the model collection on Hugging Face.

To use jina-embeddings-v5-text-small, you must have the appropriate subscription level or the trial period activated.

Create an inference endpoint that references the jina-embeddings-v5-text-small model in the model_id field.

				PUT _inference/text_embedding/eis-jina-embeddings-v5-text-small
					{
  "service": "elastic",
  "service_settings": {
    "model_id": "jina-embeddings-v5-text-small"
  }
}
		

The created inference endpoint uses the model for inference operations on Elastic Inference Service. You can reference the inference_id of the endpoint in text_embedding inference tasks or search queries. For example, the following API request ingests the input text and produce embeddings.

				POST _inference/text_embedding/eis-jina-embeddings-v5-text-small
					{
  "input": "The sky above the port was the color of television tuned to a dead channel.",
  "input_type": "ingest"
}
		

• jina-embeddings-v5-text-small works best on small, medium or large sized fields that contain natural language. For connector or web crawler use cases, this aligns best with fields like title, description, summary, or abstract. Although the model supports a 32768 token context window, consider chunking very large fields to control latency and cost.
• Larger documents take longer at ingestion time, and inference time per document also increases the more fields in a document that need to be processed.
• The more fields your pipeline has to perform inference on, the longer it takes per document to ingest.

The jina-embeddings-v5-text-nano model is a compact, multilingual dense vector embedding model that you can use through Elastic Inference Service (EIS). It is optimized for retrieval, text matching, clustering, and classification with task-specific adapters and includes support for Matryoshka Representation Learning, which enables you to truncate embeddings to fewer dimensions with minimal loss in quality. As the model runs on EIS, Elastic's own infrastructure, no ML node scaling and configuration is required to use it.

The jina-embeddings-v5-text-nano model has 239M parameters, supports an 8192 token input context window, and produces 768-dimension embeddings by default.

For more information about the model family, refer to the Elastic blog post or the model collection on Hugging Face.

To use jina-embeddings-v5-text-nano, you must have the appropriate subscription level or the trial period activated.

Create an inference endpoint that references the jina-embeddings-v5-text-nano model in the model_id field.

				PUT _inference/text_embedding/eis-jina-embeddings-v5-text-nano
					{
  "service": "elastic",
  "service_settings": {
    "model_id": "jina-embeddings-v5-text-nano"
  }
}
		

The created inference endpoint uses the model for inference operations on Elastic Inference Service. You can reference the inference_id of the endpoint in text_embedding inference tasks or search queries. For example, the following API request ingests the input text and produce embeddings.

				POST _inference/text_embedding/eis-jina-embeddings-v5-text-nano
					{
  "input": "The sky above the port was the color of television tuned to a dead channel.",
  "input_type": "ingest"
}
		

• jina-embeddings-v5-text-nano works best on small, medium or large sized fields that contain natural language. For connector or web crawler use cases, this aligns best with fields like title, description, summary, or abstract. Although the model supports an 8192 token context window, consider chunking very large fields to control latency and cost.
• Larger documents take longer at ingestion time, and inference time per document also increases the more fields in a document that need to be processed.
• The more fields your pipeline has to perform inference on, the longer it takes per document to ingest.

The jina-embeddings-v3 is a multilingual dense vector embedding model that you can use through Elastic Inference Service (EIS). It provides long-context embeddings across a wide range of languages without requiring you to configure, download, or deploy any model artifacts yourself. As the model runs on EIS, Elastic's own infrastructure, no ML node scaling and configuration is required to use it.

The jina-embeddings-v3 model supports input lengths of up to 8192 tokens and produces 1024-dimension embeddings by default. It uses task-specific adapters to optimize embeddings for different use cases (such as retrieval or classification), and includes support for Matryoshka Representation Learning, which allows you to truncate embeddings to fewer dimensions with minimal loss in quality.

For more information about the model, refer to the model card on Hugging Face.

To use jina-embeddings-v3, you must have the appropriate subscription level or the trial period activated.

Create an inference endpoint that references the jina-embeddings-v3 model in the model_id field.

				PUT _inference/text_embedding/eis-jina-embeddings-v3
					{
  "service": "elastic",
  "service_settings": {
    "model_id": "jina-embeddings-v3"
  }
}
		

The created inference endpoint uses the model for inference operations on Elastic Inference Service. You can reference the inference_id of the endpoint in text_embedding inference tasks or search queries. For example, the following API request ingests the input text and produce embeddings.

				POST _inference/text_embedding/eis-jina-embeddings-v3
					{
  "input": "The sky above the port was the color of television tuned to a dead channel.",
  "input_type": "ingest"
}
		

• jina-embeddings-v3 works best on small, medium or large sized fields that contain natural language. For connector or web crawler use cases, this aligns best with fields like title, description, summary, or abstract. Although jina-embeddings-v3 has a context window of 8192 tokens, it's best to limit the input to 2048-4096 tokens for optimal performance. For larger fields that exceed this limit - for example, body_content on web crawler documents - consider chunking the content into multiple values, where each chunk can be under 4096 tokens.
• Larger documents take longer at ingestion time, and inference time per document also increases the more fields in a document that need to be processed.
• The more fields your pipeline has to perform inference on, the longer it takes per document to ingest.

Reranker models take a query and an already retrieved set of candidate documents, then reorders those candidates by predicted relevance. Rerankers improve precision for the top query results.

jina-reranker-v3 is a multilingual listwise reranking model that improves search relevance by reordering candidate results using cross-document context. It is available out-of-the-box through Elastic Inference Service (EIS), so you can apply reranking without managing infrastructure or model resources.

The model reranks up to 64 documents together in a single inference call, which makes it a good fit for high-precision top-k reranking in hybrid search and RAG workflows.

For more information about the model, refer to the Elastic blog post or the model announcement.

To use jina-reranker-v3, you must have the appropriate subscription level or the trial period activated.

Create an inference endpoint that references the jina-reranker-v3 model in the model_id field.

				PUT _inference/rerank/eis-jina-reranker-v3
					{
  "service": "elastic",
  "service_settings": {
    "model_id": "jina-reranker-v3"
  }
}
		

The created inference endpoint uses the model for inference operations on Elastic Inference Service. You can reference the inference_id of the endpoint in rerank inference tasks. For example, the following API request ingests the input strings and ranks them by relevance:

				POST _inference/rerank/eis-jina-reranker-v3
					{
  "input": ["The Swiss Alps", "a steep hill", "a pebble", "a glacier"],
  "query": "mountain range"
}
		

jina-reranker-v3 is designed for top-k reranking and processes up to 64 candidates at a time. For larger candidate sets, rerank the most relevant results returned by your first-stage retrieval and keep your candidate list within the model's listwise limit.

jina-reranker-v2 is a multilingual cross-encoder model that helps you to improve search relevance across over 100 languages and various data types. The model significantly improves information retrieval in multilingual environments. jina-reranker-v2 is available out-of-the-box and supports Elastic deployments using the Elasticsearch Inference API. You can use the model to improve existing search applications like hybrid semantic search, retrieval augmented generation (RAG), and more. You can use the model through Elastic Inference Service (EIS), Elastic's own infrastructure, without the need of managing infrastructure and model resources.

For more information about the model, refer to the model card on Hugging Face.

To use jina-reranker-v2, you must have the appropriate subscription level or the trial period activated.

Create an inference endpoint that references the jina-reranker-v2 model in the model_id field.

				PUT _inference/rerank/eis-jina-reranker-v2
					{
  "service": "elastic",
  "service_settings": {
    "model_id": "jina-reranker-v2"
  }
}
		

The created inference endpoint uses the model for inference operations on Elastic Inference Service. You can reference the inference_id of the endpoint in rerank inference tasks. For example, the following API request ingests the input strings and ranks them by relevance:

				POST _inference/rerank/eis-jina-reranker-v2
					{
  "input": ["luke", "like", "leia", "chewy","r2d2", "star", "wars"],
  "query": "star wars main character"
}
		

jina-reranker-v2 works best on small, medium or large sized fields that contain natural language. This aligns best with fields like title, description, summary, or abstract. The model uses a context window of 1024 tokens and automatically chunks larger content. Larger documents take longer to process, and inference time also increases the more documents are present in the reranking request.

The following blog posts provide additional background and context:

• jina-embeddings-v5-omni for text, images, video, and audio
• jina-embeddings-v5-text: Compact state-of-the-art text embeddings for search and intelligent applications
• Jina rerankers bring fast, multilingual reranking to Elastic Inference Service (EIS)
• jina-embeddings-v3 is now available on Elastic Inference Service