China's OpenBMB releases "MiniCPM-OCR 1.0," a small text-recognition model; 1.3 billion parameters, reads tables and formulas too

Inference software support came first

That Chinese AI research organization OpenBMB would release "MiniCPM-OCR 1.0," a model dedicated to reading text in images, became known in advance on Oct. 22. A change adding support for the model was proposed to "vLLM," open inference software for running AI models. Support was also added to SGLang, another inference software, so the model could be used from the first day of release. @aitrackerbot, an X account that tracks the appearance of models, reported it.

OpenBMB is an organization founded by researchers from China's Tsinghua University, and it develops the "MiniCPM" series of small models. According to the proposal to vLLM, the new model has the same architecture as the existing "MiniCPM-V 4.6," which can also read images, and differs only in its type name. The weights were released five days later, on Oct. 27.

1.3 billion parameters, with tables and formulas

MiniCPM-OCR 1.0 is a small model with 1.3 billion parameters. It reads Chinese and English text, tables, mathematical formulas and set fields in documents. It outputs formulas as LaTeX and tables as HTML, and can convert entire pages into Markdown or JSON. It is licensed under Apache 2.0 and can be used commercially. In addition to vLLM and SGLang, it supports llama.cpp and Ollama, which run on local computers.

Other companies' models score higher on some benchmarks

According to the organization, the model scored 96.51 on OmniDocBench v1.6, a document-reading benchmark. On the same benchmark, China's TeleOCR scored higher at 96.91, and other companies' models also score higher on another benchmark, olmOCR-Bench. The full-page conversion score is the result of combining the model with a separate model that reads layout, and does not reflect this model alone. As of Oct. 28, no results of third parties reproducing the scores could be found.