MiniCPM-OCR 1.0 is an open-source document recognition model from OpenBMB. It turns visual document content into text and structured representations, primarily for English and Chinese. The official model card describes a 1.3-billion-parameter model with capabilities covering text, tables, formulas, charts, seals and field extraction. This article explains the model; the separate website usage guide explains our upload-to-export workflow.
Quick facts
- Model: MiniCPM-OCR 1.0, developed by OpenBMB.
- Type: a document-focused OCR and understanding model in the MiniCPM-V/o family.
- Primary languages: English and Chinese.
- License: the official model card lists Apache 2.0 for the weights, code and documentation.
- This website: an independently operated browser service at minicpmocr.com, with source/result review and Word, Markdown and JSON exports.
- Current inference provider: ModelBest, reached through our server. We do not claim to run the model weights on our own Cloudflare infrastructure.
Model facts in this article were checked against the official model card and OpenBMB repository on October 29, 2026. The website features below describe our implementation rather than every upstream capability.
What can the model recognize?
OCR extracts content from images of documents: scanned pages, photographs and screenshots. Document recognition also needs to preserve relationships between that content, such as which heading introduces a paragraph, how table cells connect and where an equation sits in the reading order.
MiniCPM-OCR is designed for these document tasks. Its upstream capabilities include text recognition, mathematical notation, table and chart content, seals and specified-field extraction. Recognized formulas can be represented as LaTeX. The official SDK also provides post-processing and document parsing tools; different tasks and pipelines produce different output representations.
The model card's capability list is not a promise that every feature is a separate button on this website. Our browser workspace currently focuses on page parsing, linked source regions, editable recognized content and document exports. We do not sell a public API or a dedicated field-extraction interface in the subscription plans.
Does the model parse a PDF by itself?
A PDF workflow needs more than a model checkpoint. A system must read or render the pages, identify document regions, recognize their content, preserve reading order and assemble the result. The upstream repository documents a full document parsing SDK and deployment options.
Our website renders PDF pages in the browser and submits them sequentially through our server for ModelBest document parsing. The result is presented beside the original page. Keep the workspace open during processing. The current website supports unlocked PDFs up to 64 MiB and 500 pages, subject to your trial or paid page allowance.
Who is this website useful for?
This service can fit a workflow when you need to:
- extract editable content from English or Chinese reports, research notes, invoices or scanned books;
- review tables and formulas against their source before using the output;
- correct a few recognition errors in the browser;
- download Word for editing, Markdown for text and LaTeX workflows, or JSON for region coordinates and reading order.
Consider another approach if your requirement is guaranteed transcription accuracy, fully offline processing, inference on your own hardware, a hosted developer API, or language coverage beyond what you have verified with your documents. Important numbers, formulas and names need human review. The example library demonstrates the interface and does not establish an accuracy benchmark.
Is this an official OpenBMB or ModelBest website?
No. Our website is independently deployed and operated. We run the application and service workflow on Cloudflare Workers, use D1 for application data, and private R2 for document storage. Accounts, page credits, subscription billing and browser exports belong to this website's service layer.
OCR inference currently uses the ModelBest API through our server. That means the website is our own deployment, while the underlying recognition computation is provided by ModelBest. Our provider credentials stay on the server. We do not claim an official affiliation, and we do not describe this service as self-hosted model inference.
How is using this website different from self-hosting?
Using the website means opening a browser, uploading a document and working with the result. You do not install model weights or configure a GPU. The service's limits, processing providers, retention rules and prices apply.
Self-hosting means running the upstream model and supporting document pipeline on infrastructure you control. It requires choosing an inference backend, provisioning suitable hardware, installing the software and maintaining the deployment. The official repository is the source for those instructions. Our developer resources link to the upstream project; purchasing a website plan does not provide a self-hosted deployment or an API key.
What happens to uploaded documents?
Selecting a file creates a local preview. Requesting recognition sends it through our server for ModelBest processing. Originals use private R2 storage with a 24-hour expiration rule; physical deletion follows the storage lifecycle schedule. Results and local recovery are available for up to 24 hours. ModelBest has its own processing and retention rules.
Review the privacy policy before sending sensitive documents. The service is not an offline OCR tool, and private storage does not mean that no external inference provider processes the content.
What does it cost to try?
One document of up to three pages can be recognized without an account. Further recognition requires a paid plan and uses one credit per page. Current launch prices are Pro $19/month or $190/year, Business $79/month or $790/year, and Enterprise $299/month or $2,990/year. Annual plans release page allowances monthly. See current pricing for page allowances and renewal terms.