MiniCPM-OCR 1.0 is a 1.3B-parameter OCR specialist model from OpenBMB that runs locally to parse scanned, photographed, bent, or tilted documents and outputs Markdown or JSON. It is not a general-purpose chat model: first check that your hardware and use case fit, then decide whether to start with the SDK, Docker, or the free API.
Match the use case first, then pick an entry point
- Scanned documents, re-photographed pages, and bent or tilted paper documents are its home turf; ordinary clean electronic documents may not need it at all.
- Code, weights, and documentation are all under the Apache 2.0 license. The weights are on Hugging Face, the code is on GitHub, and ModelScope is also available.
- The same family also includes MiniCPM-OCR Jev, a 2B model that only scores quality and does no text recognition — don't download the wrong weights.
- From day one it supports four inference frameworks: vLLM, SGLang, llama.cpp, and Ollama, with no waiting for later support.
Running it locally means preparing two environments
- Install the SDK environment: it handles layout detection and document assembly. The pip package is minicpm-ocr and the command-line tool is minicpmvocr.
- Start the inference service environment: this is where the recognition model actually runs. Keep the two environments separate instead of installing everything in one place.
- For the easy route, use the Docker image: released on November 5, 2026, it bundles the weights, the SDK, and vLLM. It requires Linux x86-64, an NVIDIA GPU, a CUDA 12.9-compatible driver, and the NVIDIA Container Toolkit.
- After a run, save the results as Markdown or JSON so they can be searched and archived later, instead of just glancing at them in the terminal.
There is also a lightweight mobile option: since November 2, 2026, the iOS app can parse documents fully offline and locally — iOS only for now, so Android users should not wait for this route.
Be clear about the limits: this model only solves document recognition. Answering questions, translating, and rewriting need separate models. The free API is fine for trying it out, but for regular batch processing, local deployment is the better choice, keeping the workflow and the data in your own hands.
