Xiaomi-OCR-0: The small and smart revolution in local document processing
If you are building document-based automations, you are likely familiar with the dilemma: on one hand, large language models (LLMs) are smart but heavy, expensive to run, and require the cloud. On the other hand, traditional OCR tools are fast but very limited in their understanding capabilities. Xiaomi is attempting to change this equation with the launch of Xiaomi-OCR-0, an ultra-compact vision-language model (VLM) that changes the rules of the game.
What is Xiaomi-OCR-0 actually?
This is a "lean" model with only 800 million parameters (0.8B), based on the Qwen3.5 architecture. Despite its modest size, it was designed specifically for document parsing and advanced OCR tasks. Unlike general-purpose models, it was trained on a massive dataset of about 170 million samples, allowing it to perform complex tasks with very high accuracy [1, 5].
Key capabilities worth knowing
The power of the model lies in its ability to not just "read" text, but to understand the structure of the document:
- •Structured extraction: The model knows how to extract text, tables (in OTSL format), and mathematical formulas (in LaTeX format) and convert them into an organized Markdown structure [4, 6].
- •Key Information Extraction (KIE): You can ask it to extract specific fields from a document and present them as clean JSON, making it a perfect tool for automating invoices, forms, and contracts [4, 7].
- •Document Question Answering (VQA): You can simply ask the model questions about the content of the document, and it will know how to answer them based on the visual and textual information within it [6, 8].
Why does this matter for your automations?
The biggest advantage here is the ability to run the model locally. Because it is small, it does not require server farms to work. You can run it on your own machine using tools like vLLM, SGLang, or even Ollama [4, 9].
In addition, the model supports MCP (Model Context Protocol) integration, which means that AI Agents can use it as a built-in "skill." This allows your agent to "look" at a document, understand it, and perform actions accordingly, all within a local and secure workflow [4, 7].
Summary and conclusions
Xiaomi-OCR-0 is proof that you don't always need a massive model to achieve impressive results. For developers and automation experts who want to build fast, cheap, and private document processing systems, this is a tool worth checking out. The weights are available for free use (under the Apache 2.0 license), and you can start experimenting with them right now via Hugging Face [2, 3, 6].
Useful links: