GitHub radar
Turn scans and photos of documents into text, tables, and formulas
MiniCPM-OCR is a pair of small models that read text in scanned or photographed documents and return it as Markdown, tables, and formulas. With them you can digitize your own documents without retyping, checking only the parts flagged as uncertain.
It's a pair of AI models for reading documents: one turns a scan or photo into text, the other scores which parts of the result you shouldn't trust.
When it helps
Useful for a one-off job digitizing Chinese or English documents — contracts, receipts, old books. Not needed for other languages or as a general-purpose assistant.
Pros
- Runs without a GPU on an ordinary laptop
- Checker model flags only uncertain parts, not the whole text
- Leads open models in several document-parsing benchmarks
- Free online API available, no installation needed
Cons
- Only reads Chinese and English text
- Bulk processing at full speed needs a GPU server
- Narrow task, not a general-purpose assistant
How to set it up — step by step
- 1Open your coding agent (Claude Code, Codex) in your project folder.
- 2Send the agent the text from the block below.
- 3If the agent asks you to pick an engine — Ollama, llama.cpp, or vLLM — choose any.
- 4Give the agent your own PDF or photo to parse into Markdown.
- 5Check the result and whatever the checker model flags as uncertain.
- 6Don't want to install anything? Try the free online API instead.
Text for your agent
Copy this and send it to your agent — Claude Code, Codex, any of them:
Deploy MiniCPM-OCR from https://github.com/OpenBMB/MiniCPM-OCR — follow the steps in SKILL.md as the authors suggest (set up the environment, download the model, parse any PDF or photographed document of mine into Markdown) and show me the result. Ask me if you need to pick an inference backend — Ollama, llama.cpp or vLLM.
Through Ollama or llama.cpp — the model is only 1.3B parameters (plus a separate 2B checker model), so it should run on a regular laptop with no GPU. For maximum speed the authors used an NVIDIA RTX 5090 with 32GB of memory — that mode is only needed for parsing documents in bulk.
Open on GitHub▌ More finds