Xiaomi OCR 0 Demo

Xiaomi OCR 0 Demo

Parses images and PDFs, extracts fields and answers questions about documents with the Xiaomi-OCR-0 model, in a Docker-based document workbench that needs a GPU.

Made by loongwayX on Hugging Face. These notes are Mindket's, written from its Space page and README; the app's own files and license are on Hugging Face.

About

What it is: Xiaomi OCR 0 Demo is a document workbench that parses images and PDFs, extracts key information and answers questions about a document. It serves the Xiaomi-OCR-0 model with vLLM on the Space's GPU, next to two separate PP-DocLayoutV3 layout workers. It runs from a Docker image on port 7860.

Who it suits: People who handle scanned or photographed documents and want text, fields and answers out of them, and engineers studying a document AI stack.

Industry | Who uses it | How they use it | Example
Finance | Back-office teams | Pull fields from invoices and forms | Upload a PDF and read the extracted fields
Legal | Paralegals | Read the text of scanned pages | Parse a scanned contract page
Operations | Records teams | Turn paper into searchable text | Process a photo of a page
Software | Engineers | Study a vLLM and layout worker setup | Read layout_server.py and start.py

People whose documents are confidential should skip the hosted Space, because uploads are archived and relayed as described below.

Where to use it:

• A browser, by visiting the Space on Hugging Face
• Windows, macOS or Linux with Docker, for a copy on your own computer
• Port 7860 on your machine, once the container runs

Use case | Who | What they get back
Parse an image or PDF | Document teams | The parsed content of the page
Extract key information | Finance and operations | Fields pulled from the document
Ask about a document | Anyone | An answer from visual question answering

How to set it up:

• Docker on Windows, macOS or Linux
• Python and pip, only to fetch the files
• A GPU, since the README recommends an A10G large with 24 GB of GPU memory, 12 vCPU and 46 GB of RAM
• A first startup that downloads the model weights

1. Install the Hugging Face hub package.
2. Download the Space with the hf command.
3. Move into the new folder.
4. Build the Docker image.
5. Run the container and publish port 7860.

pip install -U huggingface_hub
hf download loongwayX/Xiaomi-OCR-0-demo --repo-type space --local-dir Xiaomi-OCR-0-demo
cd Xiaomi-OCR-0-demo
docker build -t xiaomi-ocr-0-demo .
docker run -it -p 7860:7860 xiaomi-ocr-0-demo

How to use it: Online, visit the Space on Hugging Face and add an image or PDF to the workbench. With a local container, open port 7860 in your browser once the model and both layout workers report ready.

Variable | Default
XIAOMI_OCR_MODEL | SeerRay-Lab/Xiaomi-OCR-0
MAX_MODEL_LEN | 16384
MAX_NUM_SEQS | 4
GPU_MEMORY_UTILIZATION | 0.65
REGION_CONCURRENCY | 4
LAYOUT_CPU_THREADS | 2 per layout worker

Example: You add a PDF to the workbench and ask for its key fields. The app parses the pages with the layout workers, reads them with the model, and returns the extracted information.

For newcomers and experts: If you are new, start with one image of a printed page and ask for a plain parse. A page that is clear and upright gives the easiest first result.

If you know the field, the stack runs vLLM 0.19.1 with CUDA 12 PyTorch 2.10 and Transformers 5.8.1, while PaddlePaddle GPU 3.3.1 and PaddleX 3.7.2 live in a separate environment. Both layout workers share GPU 0 with vLLM and take requests when idle. It starts with BF16 precision, a context of about 16 thousand tokens, four sequences at a time and four region workers.

You add an image or PDF
Layout workers find the regions
Xiaomi-OCR-0 reads them through vLLM
You get text, fields or an answer

Good to know:

• Startup: the public port opens only after the model and both layout workers are ready
• Weights: they download on first startup, and caches may need to be fetched again after a rebuild
• License: apache-2
• Cost to run: a GPU such as an A10G large, no API key or paid service is named
• Data it reaches: on the hosted Space, uploads are archived locally, and a relay copies inputs and OCR results to Xiaomi FDS, with no document data sent to a Hugging Face dataset
• Not for: confidential documents on the hosted Space, because of the upload archive

On Hugging Face, 13 November 2026

• Likes: 18
• Built with: docker
• License: apache-2.0

Links

• Space: https://huggingface.co/spaces/loongwayX/Xiaomi-OCR-0-demo
• Files: https://huggingface.co/spaces/loongwayX/Xiaomi-OCR-0-demo/tree/main

Try it

Open the Space in your browser, or copy it to your own Hugging Face account from its page. Nothing in this kit runs by itself.