Xiaomi has introduced Xiaomi-OCR-0, a compact artificial intelligence model designed to recognise text, understand document layouts and extract information from complex documents. With just 0.8 billion parameters, the model aims to deliver advanced document-processing capabilities without requiring the large computing resources associated with many general-purpose AI systems.
The model is particularly interesting because it goes beyond conventional optical character recognition (OCR). It can convert document pages into structured Markdown, recognise tables and mathematical formulas, extract specific information into JSON and answer questions about document images. Xiaomi has released the model weights, technical report and implementation code, allowing developers to experiment with and deploy the technology.
Key Takeaways
- Xiaomi has released Xiaomi-OCR-0, a 0.8-billion-parameter vision-language model.
- The model is designed for document parsing, OCR, information extraction and visual question answering.
- Xiaomi reports a score of 96.83 on OmniDocBench v1.6.
- It was trained using an OCR-focused dataset containing approximately 170 million samples.
- The model can convert documents into structured Markdown, including tables and mathematical formulas.
- Developers can access the model through GitHub and Hugging Face.
- Its compact size makes local deployment more practical, although hardware requirements depend on the inference setup.
What Is Xiaomi-OCR-0?
Xiaomi-OCR-0 is a specialised vision-language model developed for understanding documents.
Traditional OCR systems primarily identify characters and convert text inside images into machine-readable text. This works well for simple printed documents but becomes more complicated when a page contains multiple columns, tables, mathematical equations, handwritten text or unusual formatting.
Xiaomi-OCR-0 attempts to solve a broader problem: understanding how information is organised on a page.
For example, when given an image of a financial report, the model can identify headings, paragraphs, tables and figures rather than simply extracting individual lines of text.
It can then produce structured output that developers can use in other applications.
The model is built on Qwen3.5-0.8B and has undergone additional training focused on document understanding.
How Xiaomi-OCR-0 Works
Xiaomi developed a training approach combining three main techniques.
1. Q-Mask text anchoring
Q-Mask is part of Xiaomi’s approach to connecting visual information with the text being recognised.
The technique helps the model associate the appearance of text within a document image with its corresponding textual representation.
This is important because document understanding requires more than recognising individual characters. The model must also understand how text relates to other elements on the page.
2. Continued pretraining
Xiaomi used an OCR-focused training corpus containing approximately 170 million samples.
The dataset combines automated data generation, expert consensus, rendered-document verification and targeted synthetic examples.
This additional training helps the model specialise in document-related tasks rather than relying only on the general knowledge acquired during its original training.
3. Mixed-task reinforcement learning
The third component is mixed-task reinforcement learning, which Xiaomi calls Mix-RL.
Instead of concentrating on a single OCR task, the training process combines several document-understanding objectives.
These include text recognition, document parsing, information extraction and visual question answering.
The goal is to produce a model that can both read documents and understand the information they contain.
Xiaomi-OCR-0 Benchmark Performance
Xiaomi reports competitive results across several document-understanding benchmarks.
The company’s technical report highlights the following scores:
| Benchmark | Xiaomi-OCR-0 score |
|---|---|
| OmniDocBench v1.6 | 96.83 |
| Real5-OmniDocBench | 95.24 |
| Wild-OmniDocBench | 87.94 |
| UniMER-Test | 97.84 |
| PubTabNet TEDS | 92.27 |
The results suggest that the model can handle different document formats, including tables, mathematical formulas and less conventional layouts.
Xiaomi also reports an average score of 83.2 across five OCR-focused visual question-answering benchmarks.
For comparison, the company lists Qwen3.5-0.8B at 72.9 and Qwen3.5-2B at 80.9 on the same five-benchmark average.
This is notable because Xiaomi-OCR-0 uses fewer parameters than the 2-billion-parameter Qwen model while reporting a higher average score.
However, benchmark scores should be interpreted carefully. They depend on evaluation methods, model configurations and datasets. Xiaomi’s results do not establish that the model will outperform every competing OCR system in every real-world application.
What Can Developers Build With Xiaomi-OCR-0?
The model supports several practical document-processing applications.
1. Convert PDFs into structured Markdown
One of its main capabilities is converting document pages into Markdown.
It can preserve important structural elements, including headings, paragraphs, tables and mathematical formulas.
This could be useful for developers building document-search systems, AI knowledge bases and automated research tools.
2. Extract information into JSON
The model supports key information extraction, allowing users to request specific fields from a document.
For example, an application could process an invoice and extract:
- Invoice number
- Customer name
- Invoice date
- Total amount
- Tax information
The extracted information can then be returned in JSON, making it easier to integrate with business applications and databases.
3. Understand tables and mathematical formulas
Tables and equations are particularly challenging for conventional OCR because their meaning depends on their structure.
Xiaomi-OCR-0 supports table recognition and mathematical formula extraction, including conversion of formulas into LaTeX.
This could help with financial reports, academic papers, scientific documents and technical manuals.
4. Answer questions about documents
The model also supports OCR-centric visual question answering.
Users can provide a document image and ask questions about its contents.
For example, someone could upload a business report and ask which quarter recorded the highest revenue or what figure appears in a particular table.
This capability could be useful in document-analysis assistants and enterprise knowledge-management systems.
Xiaomi-OCR-0 vs Traditional OCR
| Capability | Traditional OCR | Xiaomi-OCR-0 |
|---|---|---|
| Text recognition | Yes | Yes |
| Document structure | Varies by system | Supported |
| Markdown output | Usually requires additional processing | Supported |
| Table understanding | Often requires separate tools | Supported |
| Formula recognition | Limited or specialised | Supported |
| Information extraction | Usually requires additional logic | Supported |
| Questions about documents | Requires additional AI systems | Supported |
The important difference is that Xiaomi-OCR-0 combines text recognition with document understanding in one model.
This could reduce the number of separate components developers need when building document-processing applications.
Is Xiaomi-OCR-0 Open Source?
Xiaomi has published the model through its public GitHub repository and Hugging Face model page.
The GitHub repository uses the Apache 2.0 licence, making it possible for developers to inspect the code, experiment with the model and build applications under the licence’s conditions.
The project also includes an installation guide, sample documents, evaluation tools and an optional local browser demonstration.
Developers can explore the project here:
- GitHub: https://github.com/xiaomi-research/Xiaomi-OCR-0
- Hugging Face: https://huggingface.co/spaces/SeerRay-Lab/Xiaomi-OCR-0
- Technical report: https://arxiv.org/abs/2609.36136
The repository also documents an experimental Ollama setup using a community-converted BF16 GGUF model. This makes local experimentation possible without relying entirely on a cloud API.
However, developers should distinguish Xiaomi’s original model release from community-created conversions. The company has not independently validated every third-party quantised version.
The Bigger Picture
Xiaomi-OCR-0 highlights an important direction in artificial intelligence: smaller specialised models can compete effectively with larger general-purpose systems on carefully defined tasks.
Businesses process enormous quantities of invoices, contracts, research papers, financial statements and internal documents. Turning this information into structured, searchable data is a major opportunity for AI developers.
A compact model that can recognise text, understand layouts and answer questions could make document automation more accessible, particularly for organisations that want to run AI systems locally.
The release also reflects growing competition among Chinese technology companies to develop open-weight AI models for specialised applications. Rather than building only large chatbots, companies are increasingly targeting practical workflows such as coding, document analysis, image understanding and automation.
Looking Ahead
The next challenge for Xiaomi-OCR-0 will be real-world adoption. Its benchmark results are promising, but developers will need to test it on their own documents, languages, handwriting styles and scanning conditions before deploying it in production environments.
If the model performs reliably outside controlled evaluations, it could become a useful building block for document automation, enterprise search, financial data extraction and local AI assistants. Its relatively small parameter count also makes it an interesting option for developers looking to run specialised AI models without depending entirely on large cloud-based systems.
Frequently Asked Questions
1. What is Xiaomi-OCR-0?
Xiaomi-OCR-0 is a 0.8-billion-parameter vision-language model designed for OCR, document parsing, information extraction and visual question answering.
2. Is Xiaomi-OCR-0 open source?
Xiaomi has released its code under the Apache 2.0 licence and published model resources through GitHub and Hugging Face. The model is generally described as an open-weight release.
3. How large is Xiaomi-OCR-0?
The model has 0.8 billion parameters. Its actual memory requirements depend on the precision, inference framework and deployment configuration.
4. Can Xiaomi-OCR-0 read PDF documents?
Yes. Its accompanying tools support PDF and image preparation, document parsing and structured output.
5. Can Xiaomi-OCR-0 extract tables and formulas?
Yes. It supports table recognition and mathematical formula extraction, including LaTeX output.
6. Can Xiaomi-OCR-0 run locally?
Yes. Developers can install the model and use the provided tools for local inference. An experimental Ollama setup is also documented in the repository.
7. Is Xiaomi-OCR-0 better than larger AI models?
Xiaomi reports that it outperforms some larger models on selected OCR and document-understanding benchmarks. This does not mean it is better for general-purpose AI tasks or every document-processing workload.
Get the day’s top stories in your inbox
One concise email. No spam, unsubscribe anytime.
