@DataScienceDojo: A Chinese company just open-sourced an πŽπ‚π‘ that fixes something most AI-powered OCR tools quietly struggle with: the…

X AI KOLs Timeline Models

Summary

Unlimited-OCR, a new open-source OCR model from a Chinese company, solves the memory growth issue common in AI OCR tools by keeping memory usage flat regardless of document length, enabling single-pass reading of dozens of pages at 32K context. It's MIT-licensed, 3B parameters, multilingual, and already popular on GitHub.

A Chinese company just open-sourced an πŽπ‚π‘ that fixes something most AI-powered OCR tools quietly struggle with: they slow down and get heavier on memory the longer the document gets. Here's why that happens. Most of these models process a document by generating text one piece at a time, and they keep track of everything they've already read as they go. The more pages you feed in, the more that memory builds up, so a 40-page file taxes the system a lot more than a 2-page one. Unlimited-OCR changes how the model pays attention to what it's read, so that memory stays flat no matter how many pages you throw at it. That's what lets it read dozens of pages in 𝐨𝐧𝐞 𝐩𝐚𝐬𝐬, at a standard 32K context length, without splitting the document up and losing the thread between pages. - MIT licensed, 3B parameters - Multilingual out of the box - Works with Transformers and SGLang directly - Already at 6.7k stars on GitHub within days of release If you're running document parsing through Textract, Google Vision, or Azure Document Intelligence, this is worth testing locally. It's small enough to self-host, and Baidu says the same underlying technique should work for speech recognition and translation too, not just OCR. #UnlimitedOCR #DeepSeekOCR #OpenSourceAI #LLMEngineering #DocumentAI #AIEngineering
Original Article
View Cached Full Text

Cached at: 07/21/26, 08:38 AM

A Chinese company just open-sourced an πŽπ‚π‘ that fixes something most AI-powered OCR tools quietly struggle with: they slow down and get heavier on memory the longer the document gets.

Here’s why that happens. Most of these models process a document by generating text one piece at a time, and they keep track of everything they’ve already read as they go.

The more pages you feed in, the more that memory builds up, so a 40-page file taxes the system a lot more than a 2-page one.

Unlimited-OCR changes how the model pays attention to what it’s read, so that memory stays flat no matter how many pages you throw at it. That’s what lets it read dozens of pages in 𝐨𝐧𝐞 𝐩𝐚𝐬𝐬, at a standard 32K context length, without splitting the document up and losing the thread between pages.

  • MIT licensed, 3B parameters

  • Multilingual out of the box

  • Works with Transformers and SGLang directly

  • Already at 6.7k stars on GitHub within days of release

If you’re running document parsing through Textract, Google Vision, or Azure Document Intelligence, this is worth testing locally. It’s small enough to self-host, and Baidu says the same underlying technique should work for speech recognition and translation too, not just OCR.

#UnlimitedOCR #DeepSeekOCR #OpenSourceAI #LLMEngineering #DocumentAI #AIEngineering

Similar Articles

@GoSailGlobal: Current OCR processes multi-page documents page by page. Every time you turn a page, memory is reset. Today, Baidu quietly open-sourced a model on GitHub and HuggingFace called Unlimited OCR, inspired by how humans copy books: - When copying a book, you don't reread hundreds of pages every time you write a word...

X AI KOLs Timeline

Baidu has open-sourced the Unlimited OCR model, which uses a Reference Sliding Window Attention (R-SWA) mechanism to parse documents up to 32K context in a single pass, eliminating the need for page-by-page inference.

Unlimited OCR: One-Shot Long-Horizon Parsing

Hacker News Top

Baidu releases Unlimited-OCR, an open-source model for one-shot long-horizon document parsing, building upon Deepseek-OCR with support for single images, multi-page documents, and PDFs.