MiniCPM-OCR
OpenBMB/MiniCPM-OCRReviewed by AI
What it does
Two small AI models from the MiniCPM family, released with their weights. MiniCPM-OCR 1.0 (1.3B) reads Chinese and English documents and PDFs into Markdown or JSON, including tables, formulas, charts, stamps and handwriting, and copes with scans, photos, skewed and curved pages. MiniCPM-OCR Jev (2B) looks at a result and gives each region a score from 0 to 1, so you only check the parts that are likely wrong. It is for developers: a pip package, a free hosted API, and local running with vLLM, SGLang, llama.cpp or Ollama. There are demo pages to try first.
How easy it is to install
B · installer plus one simple step · Web
- Linux D
- Web B
The easiest path is the web: the maker's demo pages (a console on platform.modelbest.ai, which needs an account, and a Hugging Face Space that does not) let you drop a document in and read the result with nothing installed. Running it yourself is a Linux job: Python 3.11-3.13, an NVIDIA card with CUDA, several GB of model weights, and an installer that fetches pinned vLLM and transformers sources and compiles them (pyproject.toml, scripts/install.sh).
Numbers on GitHub
- Stars
- 721on Oct 31, 2026
- Downloads
- —all releases
- Growth
- +721stars in 4 days
Stars over time
Facts
- Runs on
- Linux, Web
- You need to install
- Python, NVIDIA graphics card
- To build from source
- Python, NVIDIA graphics card
- Disk space
- 950 MB–4.3 GB · estimate
- Accounts or keys
- None
- Cost
- Free · Code, documentation and both model weights are Apache 2.0. The maker also offers a free hosted API and demo pages; their limits are not stated in the repository.
- License
- Apache-2.0 (Permissive)
- Can replace
- FineReader, Mathpix, Acrobat
- Last update
- Oct 30, 2026
Risk signals (5)
- Attentionskill tells agent to execute a web page
README, Skill section: the instruction to give a coding agent is "Read and execute https://github.com/OpenBMB/MiniCPM-OCR/blob/main/docs/SKILL_en.md". The agent then runs setup commands on your machine from a page whose contents can change at any time; the README also offers the local copy ./docs/SKILL_en.md, which you can read first.
AI review
- Attentiondemo server without encryption
The README's "Basic Demo (CN)" link is a bare address without HTTPS: http://82.157.64.212:10111/playground/document-parse. Documents you upload there travel unencrypted and the operator of that machine is not named.
AI review
- Infodocuments leave your computer on the hosted paths
Besides local running, the README offers the demos and a "Free API" on platform.modelbest.ai / platform.modelbest.cn. Using those means uploading the document to the company's servers; the repository does not say what happens to it afterwards.
AI review
- Infobenchmark scores are the makers own
The README states "a SOTA score of 93.68 on Real5-OmniDocBench", "1.49x" the throughput of PaddleOCR-VL-1.6 and "3.62x" that of MinerU2.5 Pro. These are the authors' own measurements; a reproduction guide is included (docs/benchmark_reproduction_en.md) and the radar chart's own caption says "some values remain to be confirmed".
AI review
- Infobrand new repository
Project facts: created 2026-10-27, last push 2026-10-29, one release (sdk-v1.0.2) with no platform binaries. The README calls MiniCPM-OCR 1.0 "the first OCR-specific model in the MiniCPM-V/o family".
AI review
Evidence is shown in its original language.
What Radar could not check
- Whether the quality holds on your own documents, languages other than Chinese and English, or your handwriting
- How much graphics memory each model and backend really needs
- The real disk use of the weights and the environments (the 950 MB to 4.5 GB estimate is a guess)
- Whether the llama.cpp and Ollama route works on Windows and Mac, which the repository does not document
- The limits of the free API and what the company does with the documents sent to it
- What score threshold to use with Jev, which the README says must be checked on your own batch first