Xiaomi unveils open-source ocr model for document analysis and visual question answering

Xiaomi has released an open-source OCR model, Xiaomi-OCR-0, with 0.8 billion parameters, designed for document parsing, key information extraction, and visual question answering. The model supports Markdown and JSON output, enabling structured data extraction from invoices, receipts, and forms. It also allows cross-region retrieval, numerical calculations, and semantic reasoning over document content. The model is built on Qwen3.5-0.8B-Base and trained on 1.7 billion OCR samples. Xiaomi claims it outperforms other models in several benchmarks, including Real5-OmniDocBench and OmniDocBench v1.6. The model is integrated into Xiaomi's 'Super Xiao Ai' for real-time information extraction from screens. Developers can access the model on Hugging Face and GitHub, with additional tools for integration into existing systems.