Baidu's New Tool Reads Long Documents Without Losing Track

On June 23, 2026, Baidu released a new AI tool called Unlimited OCR that can read and understand long documents in one go. The technical details are available on arXiv, and anyone can download the code and trained model from GitHub and Hugging Face.

How Current Document Reading Works

Most document-reading AI tools today work like this: they chop a document into small pieces, process each piece separately, then try to glue the results back together. Think of it like reading one paragraph at a time with no memory of what came before — if a page three explains what an acronym stands for, by the time you reach page thirty and see that acronym again, you have forgotten the definition.

This works fine for short, simple documents. But with long PDFs full of tables, images, and footnotes, or old scanned manuscripts with messy layouts, the system often gets confused. The pieces never connect properly because the tool has no way to carry understanding from one section to the next.

The New Approach

Unlimited OCR tries something different. Instead of breaking documents into chunks, it was designed to mimic how humans actually read: you keep a running mental model of what you have seen, using it to make sense of what comes next. The tool processes an entire document in one pass — all the way through, without starting over or jumping back and forth.

The key word here is "one pass." This is not about asking the AI a few example questions first. It means the tool reads the whole document in a single, smooth operation, which matters in practice because it is faster and cheaper than other methods.

Why This Matters

Teams that work with complicated documents every day — lawyers reading contracts, banks processing loan applications, publishers extracting data from research papers — usually need tools that work well on long, complex files. Baidu has released this tool for free so anyone can test it on their own documents without waiting for a company to build an API for them.

The broader context here is that many major companies — Microsoft, Google, and others — have spent years improving document-reading AI. But most of them have stuck with the old approach of chopping documents into pieces. If Baidu's method of mimicking human memory actually works better, that would be a real shift in how the problem is tackled. The question is whether it lives up to the promise once other researchers test it independently.

What Happens Next

Baidu chose to publish both the research paper and the actual code at the same time, which is their usual approach. This means experts can start trying it out immediately and checking whether the claims hold up on standard tests used across the industry. The real measure of success will come from third-party validation, not from Baidu's own results.

For teams considering whether to use this tool, the important questions are straightforward: how well does it handle documents as long as what you need to process, does it work with languages and scripts beyond English, and what does it cost in computing power to actually run it. The technical report is the right place to start looking for those answers.