To use this website, upload an image or unlocked PDF, select Recognize document, review the linked result and choose Export. You can download Word (.docx), Markdown (.md) or structured JSON (.json). One document of up to three pages is available as a free experience without an account.
This guide describes the minicpmocr.com workspace as checked on October 29, 2026. It is not a model installation tutorial. For the model's background and deployment choices, read What is MiniCPM-OCR?.
Before you start
Prepare one supported file and check its page count:
- Images: PNG, JPEG, GIF, WebP or BMP, up to 24 MiB. Images also have a 40-megapixel and 8,192-pixel side limit.
- PDFs: an unlocked PDF, up to 64 MiB and 500 pages.
- Free experience: one document, at most three pages, checked using browser, account and network information.
- Paid use: one credit per page, with your plan's monthly allowance.
English and Chinese are the model's primary languages. A sharp scan or photograph with readable text usually gives you more useful material to review than a blurred, cropped or heavily skewed page. OCR can make mistakes; inspect important content before you use it.
1. Open the workspace and upload a file
Go to the homepage workspace. Use Try Free in the navigation, click the upload area, drag a supported file onto it, or paste an image from the clipboard. If a document is already open, use the blue Upload button at the top of the workspace to choose the next file.
Choosing the file opens a local preview. It does not start recognition. Check that you selected the correct document, confirm the page count, and use the page control to inspect a PDF if needed. File preparation or preview errors should be resolved before you request recognition.
2. Select Recognize document
The result panel shows a Recognize document button with either the free experience or the page-credit cost. Select it when you are ready to send the document for processing.
Our independently deployed application sends recognition input through its server to the ModelBest API. Provider credentials stay on the server. This is an online service; it does not run the model offline in your browser or on our own Cloudflare hardware.
For PDFs, the browser renders pages and sends them for recognition in sequence. Keep the workspace open and watch the page progress. A cancellation control is available during a running task. Failed or cancelled documents restore their allowance or credits.
If your free experience has already been used, sign in with Google or email/password and choose a subscription plan. The free allowance is for one document, rather than a free document on every visit.
3. Review the original and recognized content together
On desktop, the original document is on the left and the result is on the right. On a narrow screen, the panels stack. Entering a document or completing recognition brings the workspace controls into view. Upload, Copy all and Export remain available in the action bar while you scroll through the document workspace; the buttons become usable when their required result is ready.
To check a region, select its box in the original or its content block in the result. Linked panels locate the matching region on the other side. Use the Display menu to show recognition boxes, show reading order, hide headers and footers, or control panel linking. Source zoom and rotation controls help you inspect difficult areas.
Switch between Preview and Markdown to inspect rendered content or its markup. Use the page control to move through a multi-page document. Check reading order, merged table cells, decimal points, names and formula symbols against the original.
4. Correct recognition errors
Double-click a recognized region to edit its content. Save your change and check the rendered result again. Your edits are included in document exports.
For example, if a table value reads 1.08 but the source says 1.06, correct the relevant result block before exporting. Mathematical content uses LaTeX, so review both the expression and its rendered form. A successful task means recognition completed, not that every character is correct.
5. Export the document
When recognition succeeds, the blue Export button is available in the workspace action bar. It is also available for prepared example results. Open it and choose the format that fits your next step:
- Word (.docx): an editable document for continued writing and review. Do not assume the layout is a pixel-for-pixel reproduction of the source.
- Markdown (.md): text and document structure, including recognized tables and LaTeX content, for editors and knowledge workflows.
- Structured JSON (.json): page dimensions, content regions, labels, reading order and coordinates for downstream processing. Coordinates refer to the upright page.
An export includes all document pages and your saved edits, rather than only the page currently in view. Copy all copies the document as Markdown. You can also copy individual regions while reviewing them.
Export creates a browser download. If the file does not appear, check your browser's downloads list and any download-permission prompt. There is no need to scroll back up through a long result to find the action bar.
What if a document does not finish?
Read the workspace error message before retrying. Check the file format, size and PDF lock status. A PDF workflow also depends on the open browser tab rendering and submitting pages.
The current local file and task can be recovered for up to 24 hours. Reopening the workspace within that window may restore it. This is a recovery window, not long-term document storage; download completed results you need to keep. A failed or cancelled task restores its allowance or credits.
Understand the paid plan before continuing
Pro includes 4,000 pages per month, Business 20,000 and Enterprise 80,000. The displayed launch prices are already the actual checkout prices, at 50% of the planned standard prices. An annual subscription is paid yearly and releases page allowances monthly. Unused pages do not roll over. Subscriptions renew at the displayed launch rate until cancelled.
See pricing and billing FAQs for the current amounts and terms. The plans cover this browser service and result exports; a public developer API is not included.
Who operates the service, and where does the document go?
We independently deploy the website and application workflow on Cloudflare Workers, with D1 application data and private R2 document storage. OCR inference currently uses ModelBest's API. We are not the official OpenBMB or ModelBest website, and independently deploying the application does not mean we self-host the model weights.
Selecting a file keeps it local until you request recognition. Processing sends it through our server to ModelBest. Originals have a 24-hour R2 expiration rule, with physical deletion following the storage lifecycle schedule. Results and local recovery are available for up to 24 hours. ModelBest has its own processing and retention rules. Read the privacy policy before uploading sensitive documents.
Useful links
- Open the workspace or review the short workflow.
- What is MiniCPM-OCR?.
- Features, FAQ and pricing.
- Official model card and OpenBMB repository for upstream model information and self-hosting instructions.