Baidu releases 3.3B open source OCR model for long documents

What happened and what to do

Baidu announced Unlimited OCR 2, an open source model with 3.3 billion parameters aimed at documents of any size. According to the authors' claim, it keeps decoding latency and memory almost constant as the document grows, and they report state-of-the-art results on the MPDocBench and MarathonDocBench benchmarks. The announcement indicates an MIT license, which allows internal use and adaptation.

For a company with multi-page PDFs, contracts, reports or book archives, the practical point is whether long volumes can be processed without relying on a per-page external API. One path is to evaluate the model on real samples from the company's own archive, measuring extraction quality, time per document and memory use on available infrastructure. From there, an ingestion pipeline can convert documents into structured text, with queues, output validation and failure monitoring, feeding search, document management systems or knowledge bases for language models.

How the consultancy can help

Diagnosis of the document archive and where current OCR limits operations, followed by a comparative test of the model on real samples, measuring quality, time and memory. On-premise ingestion pipeline implementation with validation, monitoring and integration into the target system.

Next step

If your company processes long PDFs and wants to assess whether an open source OCR model can run on your infrastructure, send a short description of your case, volume and document formats. We will reply with a scoped proposal.