Unlimited OCR 2
Unlimited OCR 2 is Baidu's second open long-document parser. It attends only to the part of the page it is transcribing, so cost stays flat as documents grow.
A release of Unlimited OCR.
- Released
- October 27, 2026
- Model ID
- baidu/Unlimited-OCR-2
- Size
- 3B
- License
- MIT
- Also searched as
- unlimited ocr 2, unlimited-ocr-2, baidu/Unlimited-OCR-2
What is new in Unlimited OCR 2
- Sparse Reference Sliding Window Attention (SR-SWA) replaces the decoder's attention with a fixed set of working references, for example 512 tokens.
- Per-step latency and memory stay nearly constant however long the document is.
- At hundred-page scale, 7.0x faster per step than full attention and 5.6x faster than R-SWA, per the paper.
- State-of-the-art end-to-end results on MPDocBench and MarathonDocBench, a new hundred-page benchmark from the authors.
- Follows Unlimited OCR, released June 22, 2026. Code for the first version stays on the UnlimitedOCR1.0 branch.
Unlimited OCR 2 benchmarks
| Benchmark | Score | Note |
|---|---|---|
| Per-step speedup over full attention | 7.0x | Hundred-page documents |
| Per-step speedup over R-SWA | 5.6x | Hundred-page documents |
Scores as published by the authors. We have not rerun them.
Use Unlimited OCR 2 through an API
OCRRouter runs a different release of Unlimited OCR as baidu/unlimited-ocr.
1DeepSeek-OCROpen-source vision model that compresses pages into few tokens.1 credit / page
deepseek/deepseek-ocr1 credit / page2PaddleOCR-VLOpen-source document model from Baidu. Excellent on Chinese, Japanese and Korean.6 credits / pagepaddle/paddleocr-vl6 credits / page3GLM-OCRSmall open-source OCR model from Z.ai with layout regions. MIT licensed.6 credits / pagezai/glm-ocr6 credits / pageUnlimited OCR versions
- Unlimited OCR: all versions, price, speed and features on OCRRouter
- Unlimited OCR 2 (October 27, 2026), this page
Questions
Sources
- Unlimited-OCR-2 model card (Hugging Face)
- Unlimited OCR on GitHub
- Unlimited OCR Works 2 paper (arXiv 2610.39308)
Last updated .