MiniCPM-OCR vs LightOnOCR-3: which Apache 2.0 OCR model should you use?
MiniCPM-OCR is one 1.3B model for Chinese and English pages; LightOnOCR-3 offers 4B, 1B and 0.8B sizes with boxes and chart data. Compare tooling and fit.
Verdict. Both are Apache 2.0 OCR models from autumn 2026. MiniCPM-OCR is a single 1.3B model for Chinese and English documents, including photographed, skewed and curved pages, with an SDK, a Docker image, quantized builds and support for iOS and llama.cpp. LightOnOCR-3 is a family of 4B, 1B and 0.8B models that writes Markdown with bounding boxes and also captions figures and reads chart data, but its model card lists English only. Full-page parsing with MiniCPM-OCR depends on a separate layout model on the PaddlePaddle stack, while the 0.8B and 4B LightOnOCR-3 models need a recent Transformers release, 5.5.4 or newer. All benchmark results come from the developers.
Specifications
| Specifications | MiniCPM-OCR 1.0 | LightOnOCR-3 |
|---|---|---|
| Developer | OpenBMB | LightOn |
| Released | October 27, 2026 | October 8, 2026 |
| Total parameters | 1.3B (MiniCPM-OCR 1.0) | 4.54B (LightOnOCR-3-4B) |
| Active parameters | All | All |
| Architecture | Dense | Dense |
| Context window | 256K tokens | 256K tokens |
| Input | Text, Image | Text, Image |
| License | Apache 2.0 (Permissive) | Apache 2.0 (Permissive) |
| Memory for the smallest variant (4-bit) | 0.897 GB (MiniCPM-OCR 1.0) | 0.589 GB (LightOnOCR-3-0.8B) |
Performance claims come from the developer and have not been independently verified.
Pick MiniCPM-OCR 1.0 if
- Your documents are in Chinese or English and may be photographed, skewed or curved.
- You want to run on a phone or a Mac, since it supports iOS and llama.cpp and has quantized builds.
- You want to extract key information from invoices, forms and seals.
- You want a ready SDK and Docker image.
Pick LightOnOCR-3 if
- You want a choice of sizes, from 0.8B for speed to 4B for accuracy.
- You want bounding boxes, figure captions and chart data from the same model.
- Your documents are in English, which is the language its model card lists.
- You prefer a pipeline built on Transformers rather than PaddlePaddle.