A secure cloud AI system processes diverse files through neural networks into a searchable database.
Elastic's Jina embeddings v5 models are now in the Microsoft Foundry catalog. Elastic said on November 18 that jina-embeddings-v5-text and jina-embeddings-v5-omni are available there, deploying as a dedicated endpoint in the customer's own Microsoft Azure environment. That makes them one more embedding choice for Azure teams building retrieval-augmented generation (RAG), enterprise search or agent pipelines. Embedding models rarely get a splash, but they decide whether an AI assistant finds the right document or confidently serves the wrong one.

A secure cloud AI system processes diverse files through neural networks into a searchable database. What actually landed in the catalog​

The family has two types of models, each in two sizes. The text models are jina-embeddings-v5-text-small and jina-embeddings-v5-text-nano, which handle multilingual retrieval, matching, clustering and classification. The omni models are jina-embeddings-v5-omni-small and jina-embeddings-v5-omni-nano, which add images, audio and video in one vector space for cross-modal search with a single query.

Microsoft's embeddings catalog lists all four entries. It marks each as generally available. The Foundry catalog shows the text entries taking text input and returning embeddings. The omni entries accept text, image, audio, video and PDF input, according to the catalog listing.

Embeddings turn content into numerical vectors, so an application can find semantically related items even when the query wording differs from the indexed material. These are retrieval components. They don't generate answers themselves. They supply the "find the right stuff" half of a RAG system.

The technical specifics (text-small)​

The Foundry model card for jina-embeddings-v5-text-small gives the most detail:

  • Size: 677 million parameters. The backbone is Qwen3-0.6B-Base, distilled from Qwen3-Embedding-4B, with four task-specific LoRA adapters.
  • Vectors: 1,024 dimensions by default. They can be truncated to 768, 512, 256, 128, 64 or 32.
  • Input: up to 32,768 tokens of text.
  • Languages: trained on 32 languages and supports 93.
  • Tasks: each request picks one of four: retrieval, text matching, clustering or classification. Retrieval treats queries and passages differently.
  • Pricing: the card says only that the unit price varies by deployment type.

The card also reports average scores of 67.0 on multilingual MTEB and 71.7 on English MTEB. It says these are the highest among embedding models under 1B parameters in the model's technical report. That is a vendor claim. It isn't an independent measurement on your data.

These specs come from the text-small card. Other variants differ. Elastic's technical material lists omni-nano at 768 dimensions and an 8,192-token context. Check the card for the exact variant before sizing an index.

The "no reindexing" angle​

The Foundry card says the text-small model shares a vector space with omni-small, which produces identical text embeddings. A team could therefore add images, video, audio and PDFs to an existing text-small index without reindexing the text. That's a useful upgrade path if you start with text and expand later. It is a specific claim about that model pairing. It doesn't mean any existing index or pipeline can move over without changes.

Reading the marketing carefully​

The press release headlines "frontier-grade accuracy at lower costs." It claims the omni models outperform models up to 20x larger on image benchmarks (MIEB and ViDoRe), lead their size class on audio (MAEB), and compete with larger video-grounding models (MMEB-v2). Treat those as vendor-reported benchmark results. Some points worth weighing:

  • No price comparison. The release gives no Azure pricing, regional availability or service-level terms. The catalog says price depends on deployment type. "Lower costs" refers to compute relative to larger models. It is not a published price.
  • Benchmarks aren't your workload. Quality will depend on your data, chunking, model size, endpoint hardware and retrieval setup.
  • Known weak spots. Elastic's own technical write-up says omni performance may be weaker for finding specific videos from natural-language descriptions and for image-to-image similarity and retrieval. It describes both as being improved. A "multimodal" label doesn't mean uniform strength.
  • Task choice matters. Elastic cautions that embeddings depend on the selected task. Retrieval-oriented vectors shouldn't be assumed suitable for clustering or classification.

Why dedicated endpoints matter to IT admins​

Elastic's Han Xiao framed the pitch this way: the best embedding model is the one a team can run where its data already lives. Microsoft's Uli Homann said the models deploy as dedicated endpoints in the customer's Azure environment, helping teams build search and agent experiences within their established security and governance boundaries.

For admins, that means data doesn't have to go to a separate third-party API. Governance, networking and access control stay in your existing Azure setup. Those are the vendors' characterizations. Check the specific deployment documentation for data handling, network isolation and logging before you rely on them for regulated data.

Practical evaluation checklist​

  1. Pick the variant. Choose text-only if your content is text. Choose omni if you need a single index across text, images, audio, video and PDFs.
  2. Check the model card. Confirm dimensions, token limits and task settings for your variant.
  3. Test on your data. Build a small evaluation set with real queries and compare recall and relevance against your current embedding model.
  4. Measure cost end to end. Include endpoint charges, vector storage, latency and re-embedding effort. Smaller dimensions save storage but trade away some accuracy, as Elastic describes it.
  5. Confirm supported options. Check that Foundry exposes any truncation or quantization settings you plan to use. Don't assume it does because the model supports them.

Bottom line​

The verified news is availability. Four Jina v5 embedding models are in the Foundry catalog as generally available entries, and they deploy as dedicated endpoints in a customer's Azure environment. The benchmark and cost claims come from Elastic and need testing on your own workload. For Azure teams weighing multilingual or multimodal retrieval, that's a reasonable shortlist addition.

 

References

  1. Elastic Makes High-Accuracy, Cost-Efficient Jina AI Models Available on Microsoft Foundry - AiThority AiThority 2026-11-19T10:28:25+00:00
  2. jina-embeddings-v5-text-small | Model Catalog | Microsoft Foundry ai.azure.com
  3. Model Catalog | Microsoft Foundry ai.azure.com