Baidu releases Unlimited OCR 2 as open source for long documents

What happened and what to do

According to Baidu's announcement, Unlimited OCR 2 is an open-source document parsing model. Its new mechanism, called SR-SWA, condenses attention interactions between text and image into fixed-size working references. The company says this keeps latency and memory nearly constant as the document grows. The source describes the technique as a way to limit attention and memory in long, many-page OCR.

For a company that processes contracts, reports, medical records or digitized archives with hundreds of pages, the practical point is the cost of handling large batches without memory growing with the document. A possible implementation is an ingestion pipeline that splits documents into batches, runs parsing on the company's own infrastructure, and measures time and memory per page before choosing a production architecture. It is also worth validating extraction quality on real samples, comparing extracted fields with human review, before automating any decision about the content.

How the consultancy can help

Diagnosis of long-document volume and types, controlled testing of the model on representative samples measuring latency, memory and quality, and implementation of a parsing pipeline on the client's infrastructure, with monitoring and error review.

Next step

Send a short description of your case: document types, monthly volume, average page count and where extraction is used today. With this information we prepare a scoped diagnosis and implementation proposal.