Jina AI
Search models from Jina AI, now part of Elastic: v5 text and omni embeddings, rerankers and the Reader URL-to-Markdown API
Overview
Jina AI makes search foundation models: embeddings, rerankers and the Reader API, which turns any public URL into clean Markdown when you prefix it with r.jina.ai. Elastic completed its acquisition of the Berlin company in October 2025; Jina's former CEO Han Xiao is now Elastic's VP of AI, and Elastic says it will keep publishing Jina models on Hugging Face. The current embedding line is v5: jina-embeddings-v5-text-small (677M parameters, 32K tokens) and -nano (239M, 8K) for text, and v5-omni-small (about 1.74B, 1,024 dims, 32K) and -nano (about 1.04B, 768 dims, 8K) for text, images, audio, video and PDFs. jina-clip-v2, jina-code-embeddings (0.5B and 1.5B) and jina-reranker-v3 are also offered. You can call them through the Jina API, through Elastic Inference Service, or self-host them; most Jina weights carry a non-commercial license (CC BY-NC 4.0), so commercial self-hosting needs the Jina On-Prem license that Elastic sells. Also from Elastic: Elasticsearch, which can use Jina models for semantic search. The Reader API has its own page: Jina Reader.
The Verdict
Who Should Use Jina AI?
Best For
- Multilingual and multimodal search (text, images, audio, video, PDFs)
- Elasticsearch users who want embeddings and reranking built in
- RAG pipelines that also need web pages as clean Markdown (Reader)
- Researchers testing open weights for free
Not Ideal For
- Free commercial self-hosting: most weights are CC BY-NC 4.0
- Buyers who need a published per-token price up front
- Production use of jina-embeddings-v4 (Jina itself says it is not suited)
What's Great
- Wide model range: text, omni, code, CLIP and rerankers from one API key
- Long 32K context on the small v5 and code models
- New keys get free tokens usable on every model
- Reader API is simple: prefix a URL, get Markdown
- Rerankers now run inside Elastic Inference Service
Watch Out For
- Non-commercial licenses on most open weights (v3, code, reranker v3)
- jina-embeddings-v4 uses the Qwen Research License with no commercial option
- No per-1M-token price on the public pages; you buy token top-ups
- Fast model churn: v3 is already marked legacy, so plan for re-embedding
- Few independent reviews or community threads
Pricing
View all features & details
Embedding Models
- jina-embeddings-v5-text-small (677M, 32K) and -nano (239M, 8K), Matryoshka dims down to 32
- jina-embeddings-v5-omni-small (1,024 dims, 32K) and -nano (768 dims, 8K)
- jina-code-embeddings 0.5B and 1.5B, 32K
- jina-clip-v2 (865M, 8K), text and images
- jina-embeddings-v4 (3.8B, 32K), research license
- v1 and v2 legacy models under Apache 2.0
Other APIs
- jina-reranker-v3 (0.6B, listwise, up to 64 documents per call)
- Reader (r.jina.ai): URL to Markdown, with selector and cache headers
Where It Runs
- Jina API
- Elastic Inference Service (rerankers v2 and v3)
- Self-hosted from Hugging Face (check each model card's license)
Real-World Usage
Adoption
- 124 model repos on the jinaai Hugging Face org
- v5-omni-small shows about 1M downloads there
- The open-source Reader repo is active on GitHub (last push May 2026)
Inside Elastic
- Acquisition completed October 2025
- Jina rerankers available in Elastic Inference Service since February 2026
- Jina On-Prem sold as its own Elastic SKU since August 2026
How It Compares
| Feature | Jina AI v5 | Cohere Embed | Voyage AI | BAAI BGE |
|---|---|---|---|---|
| Modalities | Text, image, audio, video, PDF (omni) | Text, images | Text; images via multimodal model | Text; images via BGE-VL |
| Max input | 32K tokens | 128K tokens | 32K tokens | 8,192 tokens (bge-m3) |
| Open weights | Yes, mostly non-commercial | No | voyage-4-nano | Yes, MIT |
| Public price per 1M | Not listed (top-ups) | Not published | $0.02 to $0.12 | Free to self-host |
| Owner | Elastic | Cohere | MongoDB | BAAI (non-profit) |
Sources: Jina embeddings · Cohere models · Voyage pricing · bge-m3 card
Is this your tool? Add a badge