Xiaomi has open-sourced an 800 million-parameter OCR model that can read complex tables, extract information, and answer questions.

Beating AI Insight News Brief: Xiaomi has open-sourced Xiaomi-OCR-0, an 800-million-parameter model capable of three core tasks: converting text, complex tables, and mathematical formulas in documents into editable content; extracting specified fields from invoices, receipts, and certificates and outputting them in JSON format; and answering questions based on document and chart information, going beyond basic image text transcription. The model is fine-tuned on Alibaba’s Qwen3.5-0.8B, trained with roughly 170 million OCR and document understanding samples. To enhance data quality, Xiaomi had multiple OCR models cross-verify recognition results. For complex tables and formulas, the team re-renders recognition outputs into images and compares them with originals to catch issues like misalignment and missing characters. Training was conducted in phases: first mastering text recognition and layout parsing, then adding tasks including information extraction and visual question answering (VQA). In five OCR VQA tests released by Xiaomi, Xiaomi-OCR-0 notched an average score of 83.2, outperforming Qwen3.5-2B (80.9) — a model with 2.5 times more parameters — and approaching the 84.9 score of the 4B-parameter model. In the Real5 document parsing test, which simulates interferences such as scanning artifacts, paper warping, and screen photo capture, the model scored 95.24 when paired with an external layout detector. Currently, the model weights, inference code, MCP service, and Agent Skill are all open-sourced. Developers can integrate Xiaomi-OCR-0 into AI Agents to process screenshots, PDFs, and invoices, extracting data that can be directly processed by other applications.