Baidu’s free OCR model can recognize all documents simultaneously

In June 2026, Baidu made their Unlimited-OCR model available on Hugging Face. It’s a 3B parameter model that lets you extract text from images of documents (like contracts) or PDFs, all locally without using any APIs or uploading to the cloud. You can run it locally with Ollama, LM Studio, or deploy it yourself with an API that works with other AI products.

Optical Character Recognition (OCR) is the process of taking an image of some text (like a contract, or a pile of scanned invoices, or a printed document) and turning it into actual text that you can edit and search through. In the past, getting accurate OCR results on complicated documents required using a cloud service and uploading your documents to them. With Unlimited-OCR, you can do this all locally.

1What makes it different from standard OCR

In their research paper, published on arXiv in June 2026, they talk about “one-shot long-horizon parsing”, which means being able to process an entire document at once, instead of breaking it up into separate pages and then putting it back together again. This is important because it reduces errors caused by joining different parts of the document together (especially when there are tables or text that span across multiple pages). Unlimited-OCR can handle a context length of up to 32,768 tokens, which is enough to fit most multi-page contracts or reports with lots of details.

There are two resolution options: 640 pixels for fast processing, and 1024 pixels for more detailed results on complicated documents with small fonts or lots of table data. Images are first rendered from PDFs before being processed. Unlimited-OCR builds upon previous open source work such as DeepSeek-OCR and PaddleOCR, so it has similar multilingual capabilities.

2Why this matters if you build things

Much of the valuable data in your company is likely stored in PDFs: contracts, bank statements, invoices, medical reports, insurance papers, etc. Being able to extract that data and put it into a database or automate something based on it is often one of the top requests for any kind of automation project. With a free, local OCR solution like Unlimited-OCR, the economics change dramatically, especially if you have confidential data that you don’t want to send over the internet.

3 billion parameters means you don’t need fancy hardware to run it — you can run it on your personal computer’s GPU, or a small server. It’s also licensed under the MIT license, so you can use it for commercial purposes without having to pay any fees or royalties. They’ve even made quantized weights available for llama.cpp, Ollama, LM Studio, and Jan, which should cover most local use cases.