MiniCPM-OCR-1.0: a 1.3B-parameter model for document recognition released on

On October 21, 2026, the model openbmb/MiniCPM-OCR-1.0 was published on HuggingFace with the image-text-to-text task. It is the first OCR-specific model in the MiniCPM-V/o family, focused on document understanding: text, tables, formulas, charts, stamps and structured information extraction. The author claims a score of 93.68 on Real5-OmniDocBench and support for Chinese and English documents, including scans, curved pages and re-photographed images.

What it means

The model has 1.3 billion parameters, the largest weights file is model.safetensors at 2.4 GB, and the license is apache-2.0. The site hosts four quantized builds in GGUF, AWQ and GPTQ formats; the most compact working one is Q4_0 at 0.5 GB from the openbmb/MiniCPM-OCR-1.0-gguf repository. vLLM, SGLang, llama.cpp and Ollama are listed for running it, plus a variant with DSpark decoding acceleration that, according to the author, raises end-to-end processing efficiency by 55%. For running locally on your own hardware, the 0.5 GB GGUF build is of practical interest: it fits into the VRAM of even modest GPUs, and via llama.cpp or Ollama the model can also be brought up on a CPU. The full-size 2.4 GB weights in BF16 will require noticeably more resources, but they too remain within the capabilities of mid-range consumer graphics cards. More on local AI tool news is in the CPU3D news feed.