What it is
Unlimited-OCR is an open document-recognition model from Baidu, released on June 22, 2026 under the MIT licence. It starts from DeepSeek-OCR and replaces every attention layer in the decoder with what Baidu calls Reference Sliding Window Attention, which keeps memory use constant however long the output grows. The decoder is a mixture-of-experts model with 3 billion parameters, about 500 million active at a time.
The result, per Baidu's paper, is that the model can transcribe dozens of pages in one pass within a 32K-token limit; Baidu tested documents of 40 or more pages. It runs through Transformers, vLLM (with dedicated Docker images) and SGLang, can be fine-tuned with ms-swift, and has about 2.2 million downloads in the past month on Hugging Face.
Baidu AI Cloud also sells it as a hosted document-parsing API that accepts PDFs of up to 500 pages and returns Markdown or JSON. That service is free for a limited time, with 200 test pages for verified individuals and 1,000 for verified businesses.
Our review
Unlimited-OCR fixes a real cost problem in model-based OCR. Ordinary models slow down and use more memory as each page's output grows, because the attention cache keeps expanding. Baidu's change holds the cache flat, and its paper reports 5,580 tokens a second at 512 concurrent requests on OmniDocBench, against 4,951 for DeepSeek-OCR, a 12.7% gain that Baidu says widens on longer outputs.
Accuracy is strong on Baidu's numbers: 93.23 on OmniDocBench v1.5 against 89.17 for DeepSeek-OCR 2, and 93.92 on v1.6. The v1.6 lead over Baidu's own Qianfan-OCR is 0.02 points, which is noise, and PaddleOCR-VL-1.6, also from Baidu's ecosystem, reports 96.33 on v1.6 in its own card. On single pages, Unlimited-OCR is not the most accurate open model.
The long-document case rests on an in-house test. At 40-plus pages Baidu reports edit distance below 0.11, and attributes most errors to small text, because multi-page mode drops every page to 1024 by 1024 pixels. 'Unlimited' is a name rather than a property: page images still fill the 32K context, and the authors say 128K versions are the next step.
Self-hosters should read the serving example before copying it. Baidu's SGLang command binds an OpenAI-compatible server to 0.0.0.0 on port 10000 with no authentication, which on a cloud VM exposes every document you send it to anyone who finds the port.
Where it earns its place
A data engineer at a law firm digitising 30 to 50 page scanned contracts
Convert each contract to Markdown in one pass, without page-stitching code
Render the PDF to images at 300 dpi and call infer_multi; the whole document comes back as one output, so clause numbering and tables that cross pages stay intact. Small footnote text is the known weak spot at multi-page resolution.
An ML platform team at a Chinese insurer processing claim files
Run document parsing without buying GPUs while volumes are uncertain
Use Baidu AI Cloud's hosted API, free for a limited time with PDFs up to 500 pages. Submissions are capped at 2 per second and result links expire after 30 days, so downstream storage has to be built in.
A self-hosting team at a European archive with a strict no-cloud policy
OCR long reports and books on one in-house GPU
Serve it with vLLM's dedicated Docker image; with about 500M active parameters it shares a GPU comfortably. Put it behind an authenticating proxy rather than exposing the server port Baidu's example opens.
Before you commit
- Setup
- Install the pinned Transformers stack (torch 2.10, transformers 4.57.1) with trust_remote_code, or pull vllm/vllm-openai:unlimited-ocr; a CUDA 12.9 image is provided for Hopper GPUs.
- Integrations
- Transformers, vLLM, SGLang and ms-swift for fine-tuning; Baidu AI Cloud's asynchronous document-parsing API returns Markdown or JSON.
- Data and privacy
- MIT licence, no use restrictions. Self-hosted, documents stay on your hardware; on Baidu AI Cloud, result links stay valid for 30 days and wider retention terms are not stated on the API page.
- Cost at scale
- Weights are small: 3B parameters is roughly 6GB in bf16, our arithmetic, so one data-centre GPU serves it. Baidu reports 5,580 tokens a second at 512 concurrency; the GPU used is not stated in the sections we read. Baidu Cloud's paid rate is not published.
- Limits
- 32K-token context including page images; multi-page input at 1024 by 1024 per page; Baidu Cloud limits of 2 submissions and 5 result queries per second, 500 pages and 100MB per PDF.
What's good
- Constant memory use however long the output grows
- Dozens of pages per pass, tested to 40-plus
- MIT licence, with vLLM Docker images and SGLang support
- Only about 500M active parameters per token
What's not
- All benchmarks are Baidu's own; the long-document set is in-house
- Multi-page mode drops pages to 1024 by 1024, so small text suffers
- PaddleOCR-VL-1.6 reports higher single-page accuracy
- Sample SGLang server listens on all interfaces with no auth
Best for
Teams digitising long, multi-page documents who want whole-document passes on their own GPUs.
Alternatives
Unlimited-OCR: common questions
- Is Unlimited-OCR worth it?
- Unlimited-OCR turns a whole contract into one model call, and every number behind that claim so far comes from Baidu. We rate it 4.1 out of 5.
- How much does Unlimited-OCR cost?
- Free under the MIT licence. Baidu AI Cloud's hosted API is free for a limited time (200 test pages for individuals, 1,000 for businesses); paid rates not published.
- What are the downsides of Unlimited-OCR?
- All benchmarks are Baidu's own; the long-document set is in-house; multi-page mode drops pages to 1024 by 1024, so small text suffers; paddleOCR-VL-1.6 reports higher single-page accuracy; sample SGLang server listens on all interfaces with no auth.
- Who is Unlimited-OCR best for?
- Teams digitising long, multi-page documents who want whole-document passes on their own GPUs.
- How much does Unlimited-OCR cost for a team or in production?
- Weights are small: 3B parameters is roughly 6GB in bf16, our arithmetic, so one data-centre GPU serves it. Baidu reports 5,580 tokens a second at 512 concurrency; the GPU used is not stated in the sections we read. Baidu Cloud's paid rate is not published.
- Is Unlimited-OCR safe for company data?
- MIT licence, no use restrictions. Self-hosted, documents stay on your hardware; on Baidu AI Cloud, result links stay valid for 30 days and wider retention terms are not stated on the API page.
- What are the best Unlimited-OCR alternatives?
- The closest alternatives we have reviewed are PaddleOCR-VL-1.6 and DeepSeek-OCR 2.
Answers checked Oct 22, 2026.
Make Unlimited-OCR? Show this rating on your site
