Xiaomi has officially announced the open-sourcing of Xiaomi-OCR-0, a lightweight 0.8B-parameter document understanding model, releasing both the model weights and inference code. Built on Qwen3.5-0.8B-Base, the model handles three task types—document parsing, key information extraction, and OCR visual question answering—all within a single model. Training used approximately 170 million samples in total, with performance optimized through multiple rounds of data filtering, closed-loop validation, supplementation of long-tail difficult samples, and a three-stage training process. Multiple public benchmarks show that the model leads among models of the same parameter scale, with some results reaching the level of closed-source models. The technology has already been deployed in Xiaomi's Super Xiaoai on mobile devices, and the team also provides an interactive demo, MCP service, and Agent Skill toolkit for developers to integrate.