Xiaomi-OCR-0

Xiaomi-OCR-0

Parses documents and answers questions about text in images with one 0.8B OCR model from Xiaomi, covering tables, formulas and handwriting. For developers building OCR and document tools.

Made by xiaomi-research on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: Xiaomi-OCR-0 is a small vision-language model built for OCR. It parses documents, answers questions about the text in an image and pulls structured fields out of a page. It has about 0.8B parameters and starts from Qwen3.5-0.8B-Base.

Who it suits: It suits developers and teams who need document reading in their own software and want a small model they can run themselves.

Industry | Who uses it | How they use it | Example
Finance and legal | Operations teams | Parse scanned or photographed documents into text | A pile of skewed phone photos of contracts
Software | Developers | Add OCR and field extraction to an app | An upload form that reads values from a receipt
Archives and culture | Archivists | Read historical handwriting and calligraphy | Digitize old manuscripts
AI research | Researchers | Study small OCR models | Compare it with other document parsers on public benchmarks

If you only need basic text from clean screenshots, a simpler tool may serve you, and you can skip it.

Where to use it:

• Python with the transformers library
• Ollama, where the project lists an experimental BF16 build
• ModelScope, where the project lists the model
• Community GGUF and BF16 conversions listed in the card's news
• Windows, macOS or Linux, wherever Python runs

Use case | Who | What they get back
Document parsing | Developers | Page content turned into structured text
OCR-centered questions | Analysts | Answers about the text shown in an image
Key information extraction | Operations teams | Structured fields taken from a page

How to set it up: You need these first:

• Python and pip
• About 1.7 GB of disk for the BF16 weights, or about 437 MB for a 4-bit estimate

1. Install the Hugging Face hub tool.
2. Download the model files.
3. Load the model with transformers as the model card describes.

pip install -U huggingface_hub
hf download xiaomi-research/Xiaomi-OCR-0

How to use it: Give the model an image of a page and tell it what you want, either a parse of the whole document, an answer to a question, or a set of fields. The model card holds the exact way to run it.

Tool | What it does
Document parsing | Reads a page into structured text, including tables and formulas
OCR-centered question answering | Answers questions about the text in an image
Key information extraction | Pulls structured fields from a page

Example: You give it a photographed invoice and ask for the invoice number and the total. The model reads the page and returns those two fields as text rather than transcribing every line.

For newcomers and experts: OCR means reading text from pictures. This model does that and can also answer questions about what it read. A good first step is to download it and try one clean page before moving to skewed photos.

Training ran on a corpus of about 170M OCR-focused samples in three steps. Q-Mask anchoring ties text to its position, continued pretraining builds broad skills, and reinforcement learning with verifiable rewards works on hard cases. The data engine mixes expert consensus, render-guided checks, sample mining and targeted synthetic data.

A page image and a request go in
The model reads text and its layout
It parses, answers or extracts fields
Structured text comes back

Good to know: These are the authors' own scores, and the next-ranked model beats it by a small margin on one table.

Title: OmniDocBench v1.6 overall score
Xiaomi-OCR-0: 96.83
TeleOCR: 96.87
OvisOCR2: 96.58
PaddleOCR-VL-1.6: 96.33
MinerU2.5-Pro: 95.75
GLM-OCR: 95.22

• Other benchmarks: 95.24 on Real5-OmniDocBench, 87.94 on Wild-OmniDocBench, 97.84 on UniMER-Test and 92.27 on PubTabNet
• Question answering: Qwen3.5-2B scores 80.9 on average over five OCR-centered benchmarks, while this model scores 83.2
• Other versions: 10 quantized and 4 fine-tuned versions made by other people are listed
• License: Apache 2.0
• Cost to run: none named
• Data it reaches: the images and documents you give it

On Hugging Face, 13 November 2026

• Task: image text to text
• Library: transformers
• Parameters: 873M
• Weights: about 1.7 GB in BF16; about 437 MB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 29,537
• Likes: 319
• License: apache-2.0
• Based on: Qwen/Qwen3.5-0.8B-Base
• Quantized versions: 10
• Fine-tunes: 4
• Tags: transformers, safetensors, qwen3_5, image-text-to-text, document-parsing, document-understanding, visual-question-answering, key-information-extraction, vision-language-model, multimodal, xiaomi, seerray

Links

• Model: https://huggingface.co/xiaomi-research/Xiaomi-OCR-0
• Files: https://huggingface.co/xiaomi-research/Xiaomi-OCR-0/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download xiaomi-research/Xiaomi-OCR-0

Then follow the usage in its model card. Nothing in this kit runs by itself.