Run OpenBMB’s 1.3B MiniCPM-OCR 1.0 with Ollama on Windows, Mac or Linux, extract a document page image to a text file, and check it against the original.
Estimates for one page image at the default context. Not measured by YouRunAI; OpenBMB reports about 1.2 pages per second on an RTX 5090 with its accelerated vLLM pipeline, not this Ollama route.
| Machine | MiniCPM-OCR 1.0 · Q4_K_M | MiniCPM-OCR 1.0 · Q8_0 | MiniCPM-OCR 1.0 · F16 |
|---|---|---|---|
| NVIDIA · 6 GB VRAM | Fits on the GPU | Fits on the GPU | Fits on the GPU |
| AMD Radeon · 8 GB VRAM | Uses Ollama’s supported ROCm/Vulkan driver; CPU fallback otherwise | Same driver requirement | Same driver requirement |
| Intel Arc · 8 GB VRAM | Runs on CPU unless Ollama’s Vulkan backend detects the GPU; still practical | Same; still practical | Same; slower on CPU |
| Apple Silicon · 8 GB unified memory | Comfortable | Comfortable | Works; close other apps |
| CPU only · 8 GB RAM | Practical for single pages | Practical, slightly slower | Slower; prefer Q4_K_M |
Install Ollama, pull OpenBMB’s official package and turn a page image into a text file on your own computer. Then check it line by line against the page.
Select the weights you want. The instructions, downloads and settings change with your choice.
The publisher has not provided a complete memory target for this task. Check its hardware guidance before starting.
Start at 16,384 tokens of context, Ollama’s default for this package, for one page. The 1.4 GB package needs a little more memory than its size; very large page images use more context, so reduce the image to about 2,000 pixels on the long side if memory is short.
Choose where you want to run it. The instructions below update to match your selection.
Download and install the current Ollama release, then make sure it is running. On Linux the official one-line installer sets it up as a service. Use the latest version: older releases may not recognise this new model architecture.
Download OllamaIn PowerShell, pull OpenBMB’s official Q4_K_M package (1.4 GB, vision projector included). Ollama stores it for you, so there is no model folder to manage.
OpenBMB Ollama package and tagsCreate a folder named ocr-test in your home folder and save a clear image of one document page in it as page.png: a scan, a phone photo of a printed page, or a PDF page exported as PNG. MiniCPM-OCR is trained for English and Chinese documents, including tables and formulas.
In PowerShell, run the command below. It uses OpenBMB’s documented “Text Recognition:” prompt and saves the model’s output to page.md next to the image. A one-page document usually finishes in seconds once the model is loaded.
OpenBMB quick startOpen page.md beside page.png and compare a few lines, numbers and any table cells. Plain text comes back as text; OpenBMB documents formulas as LaTeX. Correct anything the model misread before relying on it. Run the same command with another image to process the next page.
page.md holds the recognised text from page.png, and spot checks of lines, numbers and table cells match the original page.
This guide covers one page image at a time through Ollama. For whole PDFs with layout, tables and reading order, use the companion guide Convert PDFs to Markdown locally with MiniCPM-OCR 1.0. Commands checked November 9, 2026; not run by YouRunAI.
Update Ollama to the latest release and pull the package again.
Keep the exact “Text Recognition:” prefix and the ./page.png path, and run the command from inside the ocr-test folder.
Keep the Out-File -Encoding utf8 part of the command and open page.md in VS Code or Notepad, not the console.
Original instructions, model files, and compatibility notes behind this setup.
Source-backed package and platform-route plan for private local chat with Qwen3.5 · 9B. Preserves the official Ollama 9B package, a smaller 4B family fallback, and a clearly labeled community uncensored GGUF option.
Run Qwen3.8 27B with Ollama, connect it to Cline in VS Code, and make one reviewable code change in a tiny local project.
Generate images from text prompts in ComfyUI with the fast distilled or more adjustable base 4B model. Save your results and reuse the visual workflow.