MiniCPM-OCR 1.0

OpenBMB's 1.3B Apache 2.0 OCR model for Chinese and English documents, with an SDK, a layout stage, and iOS and llama.cpp support.

In short. MiniCPM-OCR is a small AI model that turns photos, scans and PDFs of Chinese and English documents into Markdown, with tables and formulas. It is built for messy real-world pages such as curved or skewed photos. It runs on one consumer GPU or even a phone, and the license allows commercial use.

Summary

OpenBMB, the group behind the MiniCPM models, open-sourced MiniCPM-OCR 1.0 on 27 October 2026, together with MiniCPM-OCR Jev, a separate 2B model that scores how likely an OCR result is to be correct. MiniCPM-OCR 1.0 is the first OCR-specific model in the MiniCPM-V/o family. It has 1.3 billion parameters, pairs a SigLIP vision encoder with a Qwen3.5 language model, and was trained from MiniCPM-V 4.6. It handles Chinese and English text, tables, formulas, charts, seals and key information extraction, including scans, curved pages, photographs and recaptured documents. The BF16 checkpoint is about 2.6 GB.

Full-page parsing uses two stages. The PP-DocLayoutV3 layout model, which needs PaddlePaddle, finds the regions and the reading order, and the OCR model reads each cropped region. The minicpm-ocr SDK drives this and produces Markdown and JSON. You must first start a model service with vLLM, SGLang, llama.cpp or Ollama. OpenBMB also ships AWQ and GPTQ builds, DSpark decoding heads, an official Docker image, a free API and, since early November, an iOS demo that runs offline.

OpenBMB reports 93.68 on Real5-OmniDocBench, a score it calls state of the art, leading results on OmniDocBench v1.6, Wild-OmniDocBench and olmOCR-Bench, and 54.7 on OCRBench v2. It measured 1.2 pages per second on a GeForce RTX 5090, and says DSpark decoding improves end-to-end efficiency by 55% and that throughput is about 1.49 times that of PaddleOCR-VL-1.6 and 3.62 times that of MinerU2.5 Pro. These are the developer's own measurements. The code and the weights are under Apache 2.0.

Good for

  • Parsing Chinese and English documents, including photographed, skewed and curved pages
  • Local or on-device OCR on a consumer GPU, a Mac or a phone
  • Extracting fields from invoices, forms and seals through key information extraction
  • Production pipelines that want a ready SDK, Docker image and quantized builds

Watch out for

  • Only Chinese and English are listed as primary languages
  • Full-page parsing depends on the separate PP-DocLayoutV3 layout model and the PaddlePaddle stack
  • You must start the model service yourself before the SDK works
  • Benchmark and speed comparisons with other OCR models are the developer's own

Performance claims come from the developer and have not been independently verified.

How much memory does it need?

Estimates for the MiniCPM-OCR 1.0 variant. A mixture-of-experts model must hold all of its experts in memory even though only some are used for each token, so total parameters decide the memory and active parameters decide the speed.

Variants Parameters BF16 FP8 / INT8 4-bit (Q4) Fits on
MiniCPM-OCR 1.0 BF16 1.3B 2.99 GB 1.49 GB 0.897 GB Laptop or phone

Estimates: weights plus about 15% overhead. Real usage grows with context length and depends on the software.

VRAM calculator

The license in plain words

One of the most widely used permissive licenses. You may use, change and sell products built on the model with only light conditions.

More about Apache 2.0

See also

Frequently asked questions

What is MiniCPM-OCR 1.0?
MiniCPM-OCR is a small AI model that turns photos, scans and PDFs of Chinese and English documents into Markdown, with tables and formulas. It is built for messy real-world pages such as curved or skewed photos. It runs on one consumer GPU or even a phone, and the license allows commercial use.
Can I use MiniCPM-OCR 1.0 commercially?
Yes. MiniCPM-OCR 1.0 is released under Apache 2.0, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
Is MiniCPM-OCR 1.0 open source?
MiniCPM-OCR 1.0 is released under Apache 2.0, a license approved by the Open Source Initiative. Training data and training code may not be public, so it is best described as open weights under an open-source license.
Can I run MiniCPM-OCR 1.0 on my own computer?
Yes. The smallest variant, MiniCPM-OCR 1.0, needs about 0.897 GB of memory at 4-bit precision, which fits a modern laptop or a mid-range graphics card.
When was MiniCPM-OCR 1.0 released?
MiniCPM-OCR 1.0 was released by OpenBMB on October 27, 2026.
How long a context window does MiniCPM-OCR 1.0 have?
MiniCPM-OCR 1.0 supports a context window of up to 256K tokens.

Official links

Sources

Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.

Release date and facts last verified November 11, 2026.

Related models

Model Released Parameters Context License Smallest variant needs
PaddleOCR-VL Baidu 960M – Apache 2.0 (Permissive) 0.662 GB Laptop or phone
Xiaomi-OCR-0 Xiaomi 873M 256K Apache 2.0 (Permissive) 0.602 GB Laptop or phone
GLM-OCR Z.ai (Zhipu AI) 900M – MIT (Permissive) 0.621 GB Laptop or phone
MiniCPM-V 4.7 OpenBMB 1.55B – 35.2B 3B active 256K Apache 2.0 (Permissive) 1.07 GB Laptop or phone