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

Beating AI News Flash: Xiaomi has open-sourced Xiaomi-OCR-0, with only 800 million parameters, capable of completing three types of tasks: converting text, complex tables, and mathematical formulas in documents into editable content; extracting specified fields from invoices, receipts, and certificates and outputting JSON; and answering questions based on information in documents and charts, no longer merely transcribing the text in images.

The model is trained based on Alibaba's Qwen3.5-0.8B, using approximately 170 million OCR and document understanding samples. To improve data quality, Xiaomi had multiple OCR models cross-check recognition results. For complex tables and formulas, the recognition results are also re-rendered into images and compared with the original images to identify issues such as misalignment and missing characters. Training is also conducted in stages: first strengthening text recognition and layout parsing, then adding tasks such as information extraction and visual question answering.

In the five OCR visual question answering tests published by Xiaomi, Xiaomi-OCR-0 achieved an average score of 83.2, surpassing Qwen3.5-2B (80.9), which has 2.5 times its parameter count, and approaching the 84.9 of 4B models. In the Real5 document parsing test simulating interference such as scanning, paper bending, and screen recapture, it achieved a score of 95.24 when paired with an external layout detector.

Currently, the model weights, inference code, MCP service, and Agent Skill have all been open-sourced. Developers can integrate it into AI Agents to read screenshots, PDFs, and receipts, and extract data that can be directly handed to other programs for processing.