Baidu Open-Sources a 3B OCR Model That Reads Whole PDFs in One Pass

My notes

Summary

A short-form video claims Baidu open-sourced a small (3-billion-parameter) OCR model, described as reading an entire multi-page document (up to roughly 100 pages) in a single pass rather than chunking it page by page, preserving tables, headings, clause structure, and layout. It is described as MIT-licensed, runnable fully on local hardware, and aimed at document digitization and RAG pipeline preprocessing.

Key Insight

• The pitch is single-pass, whole-document OCR versus the typical approach of splitting PDFs into pages and losing cross-page context (e.g. tables or clauses that span page breaks).
• Claimed size is small for the capability described (3B parameters), positioned as running locally rather than requiring cloud inference, which is relevant for privacy-sensitive documents (contracts, financial statements) that should not leave a local machine.
• MIT license would make it usable commercially without the restrictions many vision-language/OCR models carry.
• This is a promotional short-form video (100x.ai account) with no link to the model repo, benchmark numbers, or paper in the captured content. Treat the specific claims (3B params, “100 pages in one shot”, accuracy on tables and formulas) as unverified marketing framing until checked against the actual release.