October 31, 2026 – Xiaomi has open-sourced a lightweight document-understanding model called Xiaomi-OCR-0, positioning the 0.8-billion-parameter model as a compact all-in-one tool for parsing documents, extracting structured fields and answering questions about visual text.
Unlike traditional OCR systems that mainly convert images to text, Xiaomi-OCR-0 is designed to understand document structure. It can parse body text, complex tables and mathematical formulas, then output structured results in Markdown or LaTeX. For invoices, receipts, ID cards and business forms, it can extract designated fields into JSON. It also supports OCR-based visual question answering, allowing users to retrieve information, perform calculations and reason over charts and documents.
According to Xiaomi’s tests, the model scored 95.24 on Real5-OmniDocBench, ranking first among compared models, and 96.83 on OmniDocBench v1.6, second overall and first among models under 1 billion parameters. Across five mainstream OCR-VQA benchmarks, it averaged 83.2 points, outperforming some larger models such as the 8-billion-parameter MiniCPM-V-4.5.
Xiaomi-OCR-0 is built on Qwen3.5-0.8B-Base and trained on about 170 million OCR and document-understanding samples. The company used multi-expert consensus, render-guided verification and mixed-task reinforcement learning to improve accuracy on tables, formulas and degraded real-world images.
The technology is already being integrated into Xiaomi phones and Super Xiaoai. Use cases include extracting pickup codes, scanning bills for smart bookkeeping, recognizing card information and creating calendar events from messages or posters. Xiaomi has released the model weights, inference code, a browser-based demo, an MCP service and an Agent Skill toolkit for developers.
