---
title: MiniCPM-OCR 1.0
emoji: 🔍
colorFrom: purple
colorTo: green
sdk: gradio
sdk_version: 6.30.0
app_file: app.py
short_description: Document image OCR — text, tables, formulas, seals, charts
python_version: "3.12"
startup_duration_timeout: 30m
---

# MiniCPM-OCR 1.0

Demo of **[openbmb/MiniCPM-OCR-1.0](https://huggingface.co/openbmb/MiniCPM-OCR-1.0)** — a 1.3B OCR-specialist
vision-language model (SigLIP vision encoder + Qwen3.5 language model) built by OpenBMB for text-centric document
understanding: text, tables, formulas, seals, charts, and key information extraction, in Chinese and English.

Upload a document crop, scan, or photo and pick a task:

| Task | Model prompt | Output |
| :--- | :--- | :--- |
| Text | `Text Recognition:` | plain text |
| Table | `Table Recognition:` | OTSL → HTML table |
| Formula | `Formula Recognition:` | LaTeX |
| Seal | `Seal Recognition:` | seal text |
| Chart | `Chart to Table:` | table |
| Custom | free-form (e.g. key info extraction) | model text |

Notes:

- Runs the `MiniCPMOcrForConditionalGeneration` implementation natively in transformers (no remote code executed
  from the checkpoint repo).
- Tables are emitted in [OTSL](https://github.com/UniModal4Reasoning/OTSL) and converted to HTML with the official
  SDK converter (ported to `otsl_to_html.py`).
- The full document-parsing pipeline in the official SDK also uses a PP-DocLayoutV3 layout model to detect regions
  and reading order; this demo calls the recognition model directly on the image you provide (whole page or crop).

## Example images

The bundled example crops come from the
[MiniCPM-OCR repo](https://github.com/OpenBMB/MiniCPM-OCR) (Apache-2.0), sourced from
[OmniDocBench v1.6](https://huggingface.co/datasets/opendatalab/OmniDocBench): a Putnam 2009 text block, a Putnam
1999 display formula, and a table from Chinese national standard GB 14536.1-1998.