小尺寸全能王！小米宣布开源0.8B OCR模型：三大核心能力拉满

Kuai Technology reported on October 31 that Xiaomi Technology recently announced that it has officially released and open sourced the lightweight all-round document understanding model Xiaomi-OCR-0, known as the "Small Size All-Purpose King". The model has only 0.8B (800 million) parameters, but integrates the three core capabilities of document parsing, key information extraction, and OCR visual question and answer, and has achieved leading results in multiple benchmark tests. Unlike traditional OCR, which mainly recognizes text, Xiaomi-OCR-0 can not only recognize text in documents, but also understand complex structures and content, achieving an upgrade from text recognition to document understanding. Specifically, Xiaomi-OCR-0 has three core capabilities: Document parsing: It can accurately identify and restore text, complex tables, and mathematical formulas, generate structured results in natural reading order, and support output in Markdown, LaTeX and other formats. Key information extraction: Specified fields can be extracted from invoices, receipts, cards and business forms, and structured data such as JSON can be directly generated. OCR visual question and answer: supports information retrieval, numerical calculation and semantic reasoning around document and chart content to achieve in-depth understanding of visual text. In terms of performance, according to the test results released by Xiaomi, Xiaomi-OCR-0 scored 95.24 points in the Real5-OmniDocBench benchmark test, ranking first among the models participating in the comparison. In the OmniDocBench v1.6 comprehensive test, this model ranked second in the overall list with a score of 96.83, and ranked first among models below 1B. In addition, in five mainstream OCR visual question and answer benchmark tests, Xiaomi-OCR-0 achieved an average score of 83.2 points, surpassing some models with larger parameter sizes, including the 8B parameter MiniCPM-V-4.5. It is understood that Xiaomi-OCR-0 is built based on Qwen3.5-0.8B-Base, and about 170 million high-quality OCR and document understanding samples were used for training. Currently, related technologies are being integrated into daily usage scenarios of Xiaomi mobile phones and Super Xiaoai, enabling meal pickup.