It's raining OCR models again! This time, OpenBMB released MiniCPM-OCR, a 1B-parameter model for document parsing I made a Hugging Face Space so you can compare its outputs against Baidu's new Unlimited OCR v2 Find the link below

@huggingface Demo:

@OpenBMB @huggingface give a shot to my fast, M-series compatible UnlimitedOCR V2 engine with a python wrapper, dynamic 4-bit quants

@OpenBMB @huggingface How is it compared to GLM OCR?

@OpenBMB @huggingface I would recommend trying both on the same inputs to compare

@OpenBMB @huggingface Great test page: that's Shakespeare's Sonnet 66, so the answer key is free. MiniCPM gets the layout right but slips on cursive words: "purest faith unhappily forsworn" comes out "proudest faith unhappily forward". A known text makes these errors easy to spot 👀

@OpenBMB @huggingface Side-by-side demos are so useful for OCR. The model that wins on a clean page may not be the one you want for a crooked scan with a messy table. Being able to compare the same document makes the choice much easier.

a side-by-side on the same document is a useful way to compare these. can the Space show reading order and table structure too, or just the extracted text?

@OpenBMB @huggingface hey @NielsRogge can you give me a good model suggestion for document layout understanding and image/graph extraction from documents?

@huggingface Awesome work, Niels! 💙 Love seeing the community build tools that make OCR models easier to explore and compare. Thanks for putting this together! 🙌

A side-by-side Space is useful, but the test set decides the result more than the model. I would run the same comparison on scanned tables, multi-column pages and low-contrast scans, and score reading order separately from character accuracy. A clean page alone will flatter both models.

@OpenBMB @huggingface which one loses reading order first on rotated scans with merged table cells, MiniCPM or Baidu?

@OpenBMB @huggingface 1B params vs unlimited v2, same page set? error rate per category or it's a demo.

@OpenBMB @huggingface A 1B model for document parsing is the small-model case in one line: narrow job, high volume, output you can check. The real test is handwriting and scanned drawings, not clean PDFs.

@OpenBMB @huggingface Side-by-side comparisons are the only way I trust OCR claims now. I once picked a model on clean demo pages and watched it fall apart on my own scanned tables, so I test messy pages first.

@OpenBMB @huggingface Do you guys have any model that can translate @xQc speech to english speech or text?

@OpenBMB @huggingface every ocr model is sota until a table spans two pages.

@OpenBMB @huggingface Side by side Spaces like this make OCR comparisons much easier to trust. Do you plan to test messier inputs like rotated scans or tables that span pages?

@huggingface @OpenBMB how does minicpm-ocr handle table structures