What is Jina OCR v1?
Jina OCR v1 is a document parser from Jina AI, which Elastic bought in October 2025, released on 14 September 2026 with a technical report on arXiv. It has 3.37 billion parameters and is a post-trained DeepSeek-OCR: the same DeepEncoder vision tower, which squeezes a 1024x1024 page view into 256 visual tokens, and the same 3B mixture-of-experts decoder, which uses about 570 million parameters per token. It outputs Markdown, with tables as HTML and formulas as LaTeX.
Jina adds two things. A FastMTP draft head proposes three tokens at a time for the main model to check, which speeds up decoding without changing the output, and post-training uses reinforcement learning with graded, code-checked rewards for formulas, tables and structure.
The weights are under CC BY-NC 4.0, so commercial self-hosting needs a separate deal with Jina's sales team. The hosted route, through Jina Reader or the Jina API, is billed per token.
How much does Jina OCR v1 cost?
| Plan | Price | What you get |
|---|---|---|
| Jina API | $2 per 1M output tokens | $0.50 per 1M input tokens |
| Weights | Free, non-commercial | CC BY-NC 4.0; commercial licence on request |
Jina API: $0.50 per 1M input tokens, $2 per 1M output tokens. Weights free for non-commercial use under CC BY-NC 4.0; commercial self-hosting needs a licence from Jina (price not published).
Is Jina OCR v1 worth it?
Jina OCR v1's case is cost per page, not accuracy. In Jina's own test on one A100 at 32 concurrent requests it parsed 2.57 pages a second, the fastest of 14 systems, mainly because it writes about 1,085 tokens per page; Surya OCR 2 decodes more tokens a second but writes 3,568 per page and finishes 1.05 pages a second. On an NVIDIA L4 the FastMTP head lifts decoding from 42.7 to 83.1 tokens a second in eager mode, but only from 158.3 to 185.6 under CUDA graphs, so the headline 'nearly double' depends on how you run vLLM.
On accuracy it sits in the middle of the field. Jina reports 91.14 on OmniDocBench v1.6, ahead of DeepSeek-OCR 2 at 90.25 but behind HunyuanOCR-1.5 at 94.74 and PaddleOCR-VL-1.6 at 96.34 in its own table. On olmOCR-Bench it scores 83.4, 7.4 points above the DeepSeek-OCR backbone, against 85.8 for Chandra OCR 2. All figures are Jina's runs; no independent leaderboard had scored it by early November 2026.
The benchmark scores were measured with a stricter prompt than the recommended default, one that drops headers, footers and figures, so a default-prompt deployment may not match them. The vLLM path needs a one-time architecture registration and a specific speculative-decoding setting, and the Transformers path ignores the draft head entirely.
The licence narrows the audience. CC BY-NC 4.0 rules out commercial self-hosting, and PaddleOCR-VL-1.6, GLM-OCR and MiniCPM-OCR 1.0 score higher under Apache 2.0 or MIT. What Jina offers instead is the hosted route: at $2 per million output tokens, a 1,085-token page costs about a fifth of a cent before input tokens, which suits teams already using Jina Reader or Elastic.
Where it earns its place
A search engineer at a company running Elasticsearch
Turn scanned PDFs into Markdown before indexing them
Call r.jina.ai with the X-Respond-With: jina-ocr-v1 header, or the chat-completions endpoint for single images. At $2 per million output tokens and about 1,085 tokens a page, 100,000 pages comes to roughly $220 of output tokens plus input.
A university NLP group building a document-parsing benchmark
Run a fast, reproducible baseline on thousands of pages
Self-host the weights under vLLM 0.21+ with the FastMTP head on an A100 or L4. Non-commercial research is what CC BY-NC 4.0 allows, and greedy verification keeps outputs identical to plain decoding.
Before you commit
- Setup
- Load from Hugging Face with trust_remote_code=True and run example.py for Transformers or vLLM; FastMTP needs vLLM 0.21 or later and register() before the engine starts. The hosted API needs only a Jina key.
- Integrations
- Jina Reader (r.jina.ai), OpenAI-compatible chat completions with streaming, a web playground, AWS, Azure and GCP listings; Transformers and vLLM for self-hosting.
- Data and privacy
- Weights CC BY-NC 4.0. Jina AI has been part of Elastic since October 2025; hosted requests go to Jina's API under its own terms.
- Cost at scale
- Hosted at $0.50 per million input and $2 per million output tokens; a page uses 256 to 1,156 visual tokens and about 1,085 output tokens. Self-hosted commercial pricing is not published.
- Limits
- 32,768-token position limit and 8,192-token output cap on the API; cold starts return HTTP 503 for 30 to 60 seconds; English and Chinese are the documented languages for element tasks.
What’s good
- 2.57 pages a second on one A100, fastest in Jina's 14-system test
- About 1,085 output tokens per page, the shortest above 83 on olmOCR-Bench
- About 570M active parameters, aimed at L4-class GPUs
- Hosted through Jina Reader and an OpenAI-compatible API
What’s not
- Weights CC BY-NC 4.0; commercial self-hosting needs a contract
- 91.14 on OmniDocBench v1.6, behind HunyuanOCR-1.5 and PaddleOCR-VL-1.6
- Benchmarks used a stricter prompt than the default
- No independent benchmark results yet
Best for
Teams already on Jina Reader or Elastic who want cheap hosted page-to-Markdown, and researchers who want a fast non-commercial baseline.
Jina OCR v1 alternatives
Jina OCR v1: common questions
- Is Jina OCR v1 worth it?
- Jina OCR v1 writes the shortest pages in its class and parses them quickly, but it is a mid-table model whose weights you cannot use commercially without a contract. We rate it 3.7 out of 5.
- How much does Jina OCR v1 cost?
- Jina API: $0.50 per 1M input tokens, $2 per 1M output tokens. Weights free for non-commercial use under CC BY-NC 4.0; commercial self-hosting needs a licence from Jina (price not published).
- What are the downsides of Jina OCR v1?
- Weights CC BY-NC 4.0; commercial self-hosting needs a contract; 91.14 on OmniDocBench v1.6, behind HunyuanOCR-1.5 and PaddleOCR-VL-1.6; benchmarks used a stricter prompt than the default; no independent benchmark results yet.
- Who is Jina OCR v1 best for?
- Teams already on Jina Reader or Elastic who want cheap hosted page-to-Markdown, and researchers who want a fast non-commercial baseline.
- How much does Jina OCR v1 cost for a team or in production?
- Hosted at $0.50 per million input and $2 per million output tokens; a page uses 256 to 1,156 visual tokens and about 1,085 output tokens. Self-hosted commercial pricing is not published.
- Is Jina OCR v1 safe for company data?
- Weights CC BY-NC 4.0. Jina AI has been part of Elastic since October 2025; hosted requests go to Jina's API under its own terms.
- What are the best Jina OCR v1 alternatives?
- The closest alternatives we have reviewed are DeepSeek-OCR 2, PaddleOCR-VL-1.6, MiniCPM-OCR 1.0, GLM-OCR and HunyuanOCR 1.5.
Answers checked Nov 2, 2026.