What it is
Jina Embeddings v5 is a family of embedding models from Jina AI, which Elastic acquired in October 2025. Embedding models turn content into vectors so software can find related material, the basis of search and retrieval-augmented generation (RAG). The text models arrived in early 2026: v5-text-small (677M parameters, 32K-token input, 1024 dimensions) and v5-text-nano (239M, 8K tokens). The omni models followed in May 2026 and add images, video, audio and PDFs: omni-small at about 1.7B parameters and omni-nano at about 1B.
For text, the omni models produce exactly the same vectors as the matching text model, so you can add other media to an existing index without rebuilding it. All models have task adapters for retrieval, classification, clustering and text matching, and vectors can be truncated to as few as 32 dimensions.
The Jina API charges $0.05 per million tokens for the small models and $0.02 for nano. Weights are on Hugging Face under the non-commercial CC BY-NC 4.0 licence; commercial self-hosting needs Elastic's Jina On-Prem licence, sold since August 2026.
Our review
Jina v5 is the small embedding family to beat on multilingual text, with a multimodal upgrade path no rival matches. Jina's model card reports 67.7 on the multilingual MMTEB benchmark and 71.7 on MTEB English v2 for v5-text-small, the highest among multilingual embedders under a billion parameters by its account. It was distilled from the much larger Qwen3-Embedding-4B, which is why a 677M model scores like a bigger one. These are Jina's own results.
The omni models are the distinctive part. Text vectors are identical to the text-only model, so a team can start with text search and add product photos, slide decks or call recordings later without re-embedding anything. Voyage offers a shared space across model sizes, but not across media. Elastic's launch post presents most multimodal comparisons as charts, so the claim that small competes with 3B to 7B models on visual documents is hard to check.
Hosted costs are low and predictable. A 600x600 image costs about 363 tokens on omni-small, so embedding 1M product photos comes to roughly $18 at $0.05 per million; video is sampled as 32 frames. Jina says it never trains on API inputs or outputs, but publishes no retention period and names no SOC 2 or ISO certification, and its own legal page warns the older terms may not reflect Elastic's current practice.
The licence is the obstacle. CC BY-NC 4.0 bars commercial use of the weights, so production use goes through the Jina API, the Elastic Inference Service, or a Jina On-Prem licence priced annually on request. Fully open rivals such as Qwen3-Embedding carry no such restriction, and at $0.05 per million tokens Jina undercuts voyage-4-large's $0.12 while offering a thinner body of independent retrieval evidence.
Where it earns its place
A retail search team with a text index of 5M product listings
Add search by photo without rebuilding the index
Keep the existing v5-text-small vectors and embed product images with omni-small into the same space. At about 363 tokens per 600x600 image, 1M photos cost roughly $18 on the Jina API.
An Elastic Cloud customer building RAG over internal wikis
Get embeddings without a separate vendor or contract
Call the preconfigured .jina-embeddings-v5-omni-small endpoint in the Elastic Inference Service; it is billed under the Elastic subscription and can truncate vectors to 32 dimensions to save storage.
A platform engineer indexing 1B tokens of multilingual support tickets
Embed the backlog once and keep query costs negligible
Indexing costs $50 on small or $20 on nano. At the paid tier's 2M tokens a minute that takes about eight hours, or about 20 minutes on Premium at 50M.
A healthcare vendor that must run search inside customer networks
Self-host embeddings with no calls out
Buy Jina On-Prem from Elastic, which runs fully offline and needs no Elasticsearch subscription; the price is annual and quoted. The free Hugging Face weights cannot be used commercially.
Before you commit
- Setup
- A new API key comes with free trial tokens; call the embeddings endpoint with a model name and task. Locally, the weights load in Transformers, sentence-transformers or vLLM 0.20.1.
- Integrations
- Jina API and Elastic Inference Service; sentence-transformers and vLLM for self-hosting. Images up to 5 MB and PDFs up to 8 MB, one PDF per request.
- Data and privacy
- Jina says it never uses API requests, inputs or outputs to train models and stores them only as needed to serve requests, under Elastic's DPA. No retention period, SOC 2 or ISO certification is named, and the API model list shows US data centres.
- Cost at scale
- 1B text tokens cost $50 on small or $20 on nano; 1M product images about $18 on omni-small. Commercial self-hosting is an annual Jina On-Prem licence at an unpublished price.
- Limits
- 32K tokens on small models, 8K on nano. Rate limits are 100 requests and 100K tokens a minute on free keys, 500 and 2M on paid, 5,000 and 50M on Premium. No batch-size limit except one PDF per request.
What's good
- Top MMTEB score under 1B parameters, per Jina
- Omni models share text vectors, so adding media needs no re-index
- Images cost about 363 tokens each on omni-small
- Hosted API from $0.02 per million tokens, never used for training
What's not
- CC BY-NC 4.0 weights; commercial self-hosting needs a quoted licence
- Multimodal results mostly published as charts
- No retention period or security certification published
Best for
Teams building multilingual or multimodal search on small models, especially inside Elastic.
Alternatives
Jina Embeddings v5: common questions
- Is Jina Embeddings v5 worth it?
- Jina v5's shared text-and-media vector space is the cleverest idea in small embeddings, and a non-commercial licence is the toll for anyone who wants to run it themselves. We rate it 4.3 out of 5.
- How much does Jina Embeddings v5 cost?
- Jina API: small models $0.05 per 1M tokens, nano $0.02. Weights free for non-commercial use; commercial self-hosting needs a paid Elastic licence.
- What are the downsides of Jina Embeddings v5?
- CC BY-NC 4.0 weights; commercial self-hosting needs a quoted licence; multimodal results mostly published as charts; no retention period or security certification published.
- Who is Jina Embeddings v5 best for?
- Teams building multilingual or multimodal search on small models, especially inside Elastic.
- How much does Jina Embeddings v5 cost for a team or in production?
- 1B text tokens cost $50 on small or $20 on nano; 1M product images about $18 on omni-small. Commercial self-hosting is an annual Jina On-Prem licence at an unpublished price.
- Is Jina Embeddings v5 safe for company data?
- Jina says it never uses API requests, inputs or outputs to train models and stores them only as needed to serve requests, under Elastic's DPA. No retention period, SOC 2 or ISO certification is named, and the API model list shows US data centres.
- What are the best Jina Embeddings v5 alternatives?
- The closest alternatives we have reviewed are Qwen3-Embedding and Qwen3-Reranker, voyage-4-large and EmbeddingGemma 2.
Answers checked Oct 22, 2026.
Make Jina Embeddings v5? Show this rating on your site
