Unlimited OCR 2: What It Is and What It Can Really Do
Unlimited OCR 2 is an open-source OCR model for long documents. Here is what is known, what is only claimed, and how to try it.
Free trial: 1 document a day (up to 3 pages, 10 MB), no account needed. The exact model is shown on every result. Paid plans coming soon. The free trial currently runs the Unlimited-OCR 2 model.
- Free trial: 1 document a day (up to 3 pages, 10 MB), no account needed.
What Is Unlimited OCR 2?
Unlimited OCR 2 was released by Baidu (GitHub, Hugging Face). It is the second version of the Unlimited OCR Works model family, published on 27 October 2026 under the MIT license. It reads page images and returns text, tables, formulas and headings in reading order. The Hugging Face model card lists about 3B parameters.
What Changed From Unlimited OCR v1
Version 1 (Unlimited-OCR, 22 June 2026, paper) added sliding-window attention (R-SWA) to a DeepSeek-OCR baseline with a 32K context. Version 2 switches to SR-SWA, which keeps a fixed set of working reference tokens, so per-step decoding cost and memory stay nearly flat as documents grow, according to the model card.
Benchmarks: Self-Reported vs Verified
The developer reports an overall score of 77.1 on MarathonDocBench, a benchmark it released at the same time, and says decoding is 7.0 times faster per step at 100 pages while accuracy holds at 300 pages (model card). These numbers are self-reported and have not been independently verified. We have not measured them ourselves.
What It Can't Do
The model returns text and Markdown, not a searchable PDF. Benchmark claims are the developer's own. Running it yourself needs a GPU; the developer tested on an NVIDIA A800-80GB.
How to Run Unlimited OCR 2 Locally
The model card documents Transformers (for quick tests) and SGLang (recommended for serving). GGUF builds need a patched llama.cpp, and stock Ollama cannot load v2 yet, according to the community GGUF model card.
Try It on LongDoc OCR
The free trial currently runs the Unlimited-OCR 2 model. Free trial: 1 document a day (up to 3 pages, 10 MB), no account needed. The exact model is shown on every result. Paid plans coming soon.
"Unlimited" is part of the model's name: its developers designed it for long documents. It does not mean LongDoc OCR offers unlimited use, and the free trial has page limits. We haven't finished testing long documents on LongDoc OCR, so we don't promise them. No GPU? Try LongDoc OCR online.
Sources
- Unlimited-OCR official GitHub repository (released by Baidu) (github.com/baidu/Unlimited-OCR)
- Unlimited-OCR-2 model card on Hugging Face (huggingface.co/baidu/Unlimited-OCR-2)
- Unlimited-OCR (v1) model card on Hugging Face (huggingface.co/baidu/Unlimited-OCR)
- Unlimited OCR Works, arXiv:2606.23050 (arxiv.org/abs/2606.23050)
- Official Hugging Face Space (Unlimited OCR Works 2 Studio) (huggingface.co/spaces/baidu/Unlimited-OCR-2)
- GGUF patch notes (community) (huggingface.co/AtomicChat/Unlimited-OCR-2-GGUF)
Frequently asked questions
Is Unlimited OCR 2 the same as Unlimited-OCR?
No. Unlimited-OCR is version 1 (June 2026); Unlimited OCR 2, also called Unlimited OCR Works 2, is the October 2026 release.
Are the benchmark scores verified?
No. They are reported by the developer and have not been independently verified.
Does LongDoc OCR offer unlimited use?
No. "Unlimited" is part of the model's name: its developers designed it for long documents. It does not mean LongDoc OCR offers unlimited use, and the free trial has page limits. We haven't finished testing long documents on LongDoc OCR, so we don't promise them.
Which model does LongDoc OCR run?
The free trial currently runs the Unlimited-OCR 2 model. The exact model is shown on every result. If OCR fails, you get an error message and your free use is restored; we do not switch to another engine.
Can I use the model commercially?
The model is published under the MIT license. Use of LongDoc OCR is covered by our Terms.