OCR models / OpenBMB
MiniCPM-V 4.6
OpenBMB · released 2026-04-13 · data checked 2026-11-06
MiniCPM-V 4.6 is a 1.3B-parameter general vision-language model from OpenBMB (CN), released 2026-04-13 under the Apache-2.0 license. It scores 86.93 on OmniDocBench v1.6 (#31 of 33). It needs about 3.6 GB of memory at 16-bit.
A general on-device vision-language model; MiniCPM-OCR 1.0 is its OCR fine-tune. The v1.6 score comes from the MiniCPM-V 4.7 chart; the 4.6 card itself reports 84.6 on OmniDocBench v1.5.
OmniDocBench v1.6
86.93
#31 of 33 models
olmOCR-Bench
—
Not reported
Parameters
1.3B
General vision-language model
Memory to run
~3.6 GB
16-bit weights plus about 1 GB
Benchmarks
OmniDocBench v1.6 Lab-reported86.93
Read from the chart image on the model card (4x visual token compression). Best: OXR 1.0 97.03
The tick marks the best score on each benchmark. See how to read the ranking.
Specifications
- Lab
- OpenBMB (CN)
- Type
- General vision-language model
- Total parameters
- 1.3B
- License
- Apache-2.0
- Memory to run
- ~3.6 GB at 16-bit
- Released
- 2026-04-13
Closest alternatives
- Dolphin 1.5 (400M, 86.52 on OmniDocBench, MIT)
- MiniCPM-V 4.7 1B (1.6B, 85.39 on OmniDocBench, Apache-2.0)
- MonkeyOCR-pro 3B (3.8B, 88.57 on OmniDocBench, Non-commercial)
- OCRVerse (4.8B, 88.60 on OmniDocBench, Not stated)
Sources
- https://huggingface.co/openbmb/MiniCPM-V-4.6
- https://raw.githubusercontent.com/OpenBMB/MiniCPM-o/main/assets/minicpmv4.7/overall_performance_1b_instruct.png
Something wrong here? Report it.