# Unlimited OCR

> **TL;DR.** Unlimited OCR is a 3-billion-parameter open-weight model from Baidu that transcribes multi-page documents in a single inference pass — eliminating the page-by-page chunking that makes traditional OCR pipelines brittle.

- **Category:** AI / Document Intelligence / OCR
- **Stage:** nascent
- **Age:** 3 days
- **Origin date:** 2026-06-22
- **First detected:** 2026-06-24
- **Canonical URL:** https://earlyterms.com/term/unlimited-ocr
- **Sources:** 6 primary URLs

## Definition

Unlimited OCR is a 3-billion-parameter open-weight model from Baidu that transcribes multi-page documents in a single inference pass — eliminating the page-by-page chunking that makes traditional OCR pipelines brittle. The core innovation is Reference Sliding Window Attention (R-SWA), which holds KV cache at constant size regardless of output length.

Baidu released the model on [June 22, 2026](https://github.com/baidu/Unlimited-OCR) under the MIT license alongside the arXiv paper "Unlimited OCR Works: Welcome the Era of One-shot Long-horizon Parsing." The model achieves 93.92% on OmniDocBench v1.6 — a 6-point gain over DeepSeek OCR — and processes 40+ page PDFs in a single forward pass within a 32K token context window.

## Example

A legal team processing 50-page contracts can run `baidu/Unlimited-OCR` locally via Ollama — the model ingests the full PDF image in one pass, extracts tables, formulas, and dense text with consistent layout awareness, and outputs structured Markdown. No page boundaries to stitch, no context loss mid-document.

## Analogy

Think of it as a photographic memory for scanners — one look at the whole stack, then write it all out.

## Why it's emerging now

Baidu released Unlimited OCR on June 22, 2026, solving the KV cache blowup that forced every long-document OCR pipeline to chunk by page. The MIT license and Ollama/vLLM compatibility mean teams can swap it in without a managed API, and the 93.92% OmniDocBench v1.6 score beats DeepSeek OCR by over 6 points.

## Related terms

- *competitor:* DeepSeek OCR
- *competitor:* Mistral OCR
- *related:* PaddleOCR
- *related:* OmniDocBench
- *child:* Reference Sliding Window Attention
- *parent:* Document Intelligence
- *parent:* vision-language model
- *related:* vLLM
- *related:* SGLang

## Sources

1. [baidu/Unlimited-OCR — official GitHub repository](https://github.com/baidu/Unlimited-OCR)
2. [Unlimited OCR Works — arXiv paper (Jun 22, 2026)](https://arxiv.org/abs/2606.23050)
3. [baidu/Unlimited-OCR — Hugging Face model card](https://huggingface.co/baidu/Unlimited-OCR)
4. [Hacker News — Unlimited OCR: One-shot long-horizon parsing (478 pts)](https://news.ycombinator.com/item?id=48643426)
5. [AI Weekly — Baidu Releases MIT-Licensed 3B OCR Model for Long Documents](https://aiweekly.co/alerts/baidu-releases-mit-licensed-3b-ocr-model-for-long-documents)
6. [Data Science in Your Pocket — Baidu's Unlimited OCR: Beats DeepSeek OCR, Parses Entire Book in One Go](https://medium.com/data-science-in-your-pocket/baidus-unlimited-ocr-beats-deepseek-ocr-parses-entire-book-in-one-go-6e3e1a8c9b34)

---
_Generated by EarlyTerms · https://earlyterms.com/term/unlimited-ocr_