The Avocado Pit (TL;DR)
- 🥑 Baidu's Unlimited OCR 2 can handle hundred-page documents in one go.
- 🥑 Scored a juicy 77.1 on MarathonDocBench with SR-SWA attention.
- 🥑 MIT-licensed and ready to rock your document parsing world for free.
Introduction
Welcome to the 2026 tech scene, where Baidu is squeezing the juice out of document parsing with its freshly minted Unlimited OCR 2. If you've ever been bogged down by a novel-length PDF, this is your moment of zen. Just think of it as hiring a super-efficient intern who actually reads the whole document.
Why It Matters
Baidu's open-sourcing of Unlimited OCR 2 is like handing out free golden tickets to the document processing Willy Wonka factory. This isn't just about speed—it's about making the once-daunting task of parsing hundred-page documents as easy as pie. The model's ability to keep per-step decoding costs nearly flat is a game-changer in AI efficiency, setting a new standard in the marathon of document processing.
What This Means for You
For the curious beginner or the seasoned techie, this means your document-heavy tasks just got a whole lot lighter. Whether you're a researcher sifting through endless reports or a company looking to streamline your data processing, Unlimited OCR 2 is your new best friend. And did I mention it's MIT-licensed? That means it's as free as a bird—and who doesn't love free stuff?
The Source Code (Summary)
Baidu has released Unlimited OCR 2, a 3B model that can parse hundred-page documents in a single request. It uses SR-SWA attention to maintain nearly flat per-step decoding costs, scoring an impressive 77.1 on the MarathonDocBench. This development, highlighted by MarkTechPost, positions Baidu at the forefront of long document OCR capabilities.
Fresh Take
In the grand tapestry of AI advancements, Baidu's Unlimited OCR 2 is a particularly bright thread. While many models struggle with long documents like a confused intern on their first day, Unlimited OCR 2 breezes through them. This isn't just a step forward—it's a full-on sprint. As we move further into the AI-driven future, tools like this will continue to redefine our relationship with information, making the once-impossible not just possible, but downright easy.
Read the full MarkTechPost article → Click here
