Xiaomi Releases OCR-0 Open Model for Document Recognition

Xiaomi has published OCR-0, a model designed to recognize text, extract required fields from documents, and answer content-related questions. The new model comes with a compact size of 0.8 billion parameters and is fully open-source.

• The model is built on Qwen3.5-0.8B-Base and trained on a corpus of 170 million examples.
• It supports text, tables, mathematical formulas, and exports results in Markdown and LaTeX formats.
• OCR-0 scored 95.24 and 96.83 on Real5-OmniDocBench and OmniDocBench v1.6 benchmarks respectively.
• Source code and model weights are publicly available under the Apache 2.0 license.

Architecture and Training Process

At the core of OCR-0 is the Qwen3.5-0.8B-Base model, which has been fine-tuned on a massive corpus of 170 million examples. The dataset includes full document pages, individual document fragments, and reading comprehension tasks.

Following the pre-training stage, the developers applied reinforcement learning techniques on complex examples. This approach significantly enhanced the model's accuracy in data extraction and contextual understanding.

Capabilities and Output Formats

Xiaomi-OCR-0 is capable of not only recognizing plain text but also processing complex elements such as tables and mathematical formulas. The system accurately preserves the reading order of pages, which is essential for proper data structuring.

Data can be extracted using JSON schemas, while outputs can be formatted in Markdown for text and LaTeX for mathematical formulas. This makes integration into various applications smooth and efficient.

Benchmark Results

During testing, OCR-0 scored 95.24 and 96.83 points on the Real5-OmniDocBench and OmniDocBench v1.6 benchmarks respectively. The average score across five different tests stands at 83.2 points.

For comparison, the larger Qwen3.5-2B model achieves a score of 80.9 points in similar tests. This demonstrates the high efficiency of the OCR-0 architecture despite its smaller parameter count.

Availability and Integration Tools

The code and weights of OCR-0 are distributed under the Apache 2.0 open-source license. The official repository includes tools for running the model locally.

Additionally, the repository features an MCP server and a skill for connecting to AI agents. An interactive demo is also available on Hugging Face for quick testing without installation.