Transformers 5.21 adds MiniCPM-OCR support for text-heavy document images

[Transformers 5.21](https://github.com/huggingface/transformers/releases/tag/v5.21), released October 26, adds MiniCPM‑OCR to the library’s supported models. The release describes it as an OpenBMB document‑understanding model for text‑heavy images, including scanned pages, screenshots, receipts and tables. The development is support inside Transformers, rather than an announcement that the underlying model first appeared on October 27. For developers building document workflows around the library, that creates a documented way to connect an image and a transcription request to MiniCPM‑OCR. It does not establish how accurately the model will read a particular document, or how quickly it will run on a team’s hardware.

The [MiniCPM‑OCR documentation](https://huggingface.co/docs/transformers/main/en/model_doc/minicpm_ocr) illustrates the `image‑text‑to‑text` pipeline with the checkpoint identifier `openbmb/MiniCPM‑OCR‑1.0`. Its example supplies an image in a user message, asks for the document’s text and returns generated text. This is a multimodal generation workflow, rather than a guarantee that every page becomes a complete structured record. The same documentation describes `MiniCPMOcrProcessor` as a wrapper around image processing, video processing and tokenization. For an image workflow, the processor and model configuration are part of preparing the visual input and its accompanying text. The library release supplies the integration; an application still has to decide which document to submit and how to handle the result. There is a version boundary worth preserving. The captured documentation is the development `main` branch, which explicitly requires installation from source, and it points regular package users toward the stable 5.21.0 documentation. A developer should match examples and settings to the installed version. OpenTools has not executed the pipeline or verified installation in a local environment.

The [official model card](https://huggingface.co/openbmb/MiniCPM‑OCR‑1.0) separates the model’s recognition work from its full‑document SDK. In that SDK workflow, PDF pages are rasterized, a PP‑DocLayoutV3 component finds regions and their reading order, and cropped regions are sent for recognition. Results are then assembled into Markdown; the card describes tables being represented as HTML and formulas as LaTeX. That distinction matters when choosing an integration. Sending a receipt image to the Transformers example is a narrower job than organizing a multi‑page PDF into correctly ordered regions and structured output. Library support alone should not be presented as proof that an entire PDF‑processing system, including layout and assembly, has been installed or validated.

The [model card](https://huggingface.co/openbmb/MiniCPM‑OCR‑1.0) states that the weights, code and documentation are released under Apache License 2.0. Its captured Inference Providers panel lists no provider deployment. Neither the release nor that listing establishes a newly available hosted OCR service or a hardware minimum for a particular workload. The card also warns that requested field extraction produces model‑generated text, and tells users to parse the JSON and check its fields and values before downstream use. For a document application, keeping that validation step explicit is more useful than assuming that a returned string is a verified amount, name or table cell. The supported news is the new Transformers integration and its documented loading route; accuracy and throughput need evidence from the actual documents and environment being used. An Epson Perfection V850 Pro flatbed scanner, photographed in 2023. Context for creating document images, not a MiniCPM‑OCR device requirement or test result. Photo: [Santeri Viinamäki](https://commons.wikimedia.org/wiki/User:Zunter) / [Wikimedia Commons](https://commons.wikimedia.org/wiki/File:Epson_V850_scanner_open_20230920.jpg), [CC BY‑SA 4.0](https://creativecommons.org/licenses/by‑sa/4.0/). Native crop and downscale; image adaptations retain CC BY‑SA 4.0. No artificial upscaling.