Blog

4 modalities, 100+ languages, 1 vector space: Jina Embeddings v5 in Microsoft Foundry

Run the models on managed compute in your own Azure subscription, billed by compute hour, where nano indexes 3.4x faster than the more accurate small models.

Get hands-on with Elasticsearch: Dive into our sample notebooks in the Elasticsearch Labs repo, start a free cloud trial, or try Elastic on your local machine now.

One query can return a video clip and a paragraph in the same result set. Jina Embeddings v5 is now in the Microsoft Foundry model catalog, with four models between 239 million and 1.74 billion parameters. The text models cover more than 100 languages, and the omni models put text, images, PDFs, audio, and video into one vector space. 

You deploy them as dedicated endpoints in your own Azure subscription and pay by compute hour; the model and your data stay in your tenant. The small models score 58.00 on the Massive Image Embedding Benchmark (MIEB) and 67.00 on the Massive Multilingual Text Embedding Benchmark (MMTEB); nano gives up about a point and a half on text and indexes 3.4x faster. Jina AI is part of Elastic, so the vectors go straight into Elasticsearch.

What’s in the Jina Embeddings v5 model family

Jina AI provides small, purpose-built models that preserve meaning across the entire search and retrieval process, delivering frontier-grade accuracy without the high costs. Jina models rank at or near the top of benchmarks, despite being a fraction of the size of their peers. In Foundry, Jina Embeddings v5 models bring highly relevant semantic search at lower compute costs. For full results, see the text and omni technical reports.

The Jina Embeddings v5 family includes two types of models, each available in two parameter sizes:

  • Text models (jina-embeddings-v5-text-small, jina-embeddings-v5-text-nano). Multilingual text embeddings for retrieval, matching, clustering, and classification.

  • Omni models (jina-embeddings-v5-omni-small, jina-embeddings-v5-omni-nano). Add support for images, audio, and video, all in one vector space, enabling unified cross-modal search with a single query.

Jina Embeddings v5 text models bring high accuracy in a small footprint across more than 100 languages. Omni models embed all modalities in the same vector space so a single query can return highly relevant results across all covered media and languages.

How the small and nano models compare on accuracy and speed

Choose the small models when accuracy is your most important requirement. They have larger context windows at 32,000 tokens and more parameters to maximize the relevance of the content that they embed and retrieve. 

When speed and throughput are more important, choose the nano models. The following benchmarks illustrate this tradeoff. 

For multimodal applications, v5-omni-small scores 58.00 on MIEB (image) and 41.20 on the Massive Multimodal Embedding Benchmark–Video (MMEB-Video) subset; whereas v5-omni-nano scores 47.87 and 26.87, respectively.  

Both models have similar performance for text (small scores 67.00 on MMTEB; and nano scores 65.5 on MMTEB) and audio (both score 49.69 on the Massive Audio Embedding Benchmark [MAEB]). For throughput benchmarks, we see nano models processing documents 3.4x faster for text, 2.2x faster for images and video, and 1.5x faster for audio than small models.

Model

Input

Parameters

Input token length

Matryoshka dimensions

Best for

jina-embeddings-v5-text-small

Text

677 million

32,000

1024–32

Highest semantic search accuracy in a small footprint

jina-embeddings-v5-text-nano

Text

239 million

8,000

768–32

Up to 3.4x faster document indexing and retrieval in a tiny footprint

jina-embeddings-v5-omni-small

Text, image, PDF, audio, video

~1.74 billion 

32,000

1024–32

Highest semantic search accuracy, incorporating text, images, video, and audio

jina-embeddings-v5-omni-nano

Text, image, PDF, audio, video

~1.04 billion 

8,000

768–32

1.5–3.4x faster indexing and retrieval of multimodal content

Table 1. Differences between the variants of the jina-embeddings-v5 model family.

How to deploy Jina Embeddings v5 in Microsoft Foundry

Provision a Jina Embeddings v5 endpoint on managed compute

  1. Log in to your Azure account. 

  2. Open the Microsoft Foundry model catalog, and locate Jina Embeddings v5.

  3. Select your model variant, and deploy it to managed compute. Choose a supported GPU SKU (the model cards list A10), and make sure your Azure subscription has the required quota for that SKU. Set the number of instances. (Each one runs its own copy of the model, so add instances for more traffic.) 

  4. Select Deploy.

  5. Your endpoint provisions in 15 to 20 minutes.

Call your Jina Embeddings v5 endpoint

In Models + endpoints, open your deployment and copy the Target URI and key. Use the URI as shown; it already ends in /v1/embeddings. The request body is the one that the Jina Embeddings API uses:

curl "$TARGET_URI" \
  -H "Authorization: Bearer $ENDPOINT_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "jina-embeddings-v5-text-small",
    "task": "retrieval.passage",
    "input": ["Invoices are due 30 days after delivery."]
  }'

When you call it:

  • Set task to retrieval.passage when you index and to retrieval.query when you search.

  • With an omni model, pass media in the same list as text, as Base64; for example, {"image": "<base64-encoded image>"}.

  • Catalog deployments block outbound access, so media URLs aren’t supported. 

Then store the vectors in Elasticsearch or another vector store.

Jina Embeddings v5 in the Microsoft Foundry model catalog

For more information about the Jina models in the Microsoft Foundry model catalog, click the following links. Each model has its own catalog listing, with the model card, supported GPU SKUs, and pricing:

Jina Embeddings v5 model cards and technical reports

How helpful was this content?

Related Content

Too much of a good thing: Why 20 candidates beat 200 when reranking search results

Too much of a good thing: Why 20 candidates beat 200 when reranking search results

Scott Martens
56 scanned pages to a cited answer in 5 minutes: RAG pipeline for the documents search cannot read

56 scanned pages to a cited answer in 5 minutes: RAG pipeline for the documents search cannot read

Ben-Amin York Jr.
AI video search with Elasticsearch and Jina: Find the exact seconds of footage you need

AI video search with Elasticsearch and Jina: Find the exact seconds of footage you need

JD Armada
jina-ocr-v1: One OCR model for layout, tables, math and 100+ languages

jina-ocr-v1: One OCR model for layout, tables, math and 100+ languages

Scott Martens
0.35% trained, 100% competitive: the frozen-tower architecture behind jina-embeddings-v5-omni

0.35% trained, 100% competitive: the frozen-tower architecture behind jina-embeddings-v5-omni

Jon Avezbaki

Ready to build state of the art search experiences?

Sufficiently advanced search isn’t achieved with the efforts of one. Elasticsearch is powered by data scientists, ML ops, engineers, and many more who are just as passionate about search as you are. Let’s connect and work together to build the magical search experience that will get you the results you want.