Xiaomi open-sources 0.8B Xiaomi-OCR-0 document understanding model
Xiaomi released the open-source Xiaomi-OCR-0, a 0.8B-parameter model that combines document parsing, key-information extraction, and OCR visual question answering. According to IT之家 and 智东西, it can recognize text, tables, and mathematical formulas, emit Markdown, LaTeX, and JSON outputs, and is available with inference code, a browser demo, MCP service, and an Agent Skill toolkit. Xiaomi’s reported results include 96.83 on OmniDocBench v1.6, ranking first among models below 1B parameters and second overall in the cited comparison, while the benchmark claims are vendor-reported.
Xiaomi-OCR-0 packages several document-understanding tasks into a lightweight model and publishes both weights and inference code, giving developers an accessible option for device-side document automation and agent integrations.
IT之家 10 月 30 日消息，小米官方今日宣布开源轻量级文档理解模型 Xiaomi-OCR-0，模型权重与推理代码已同步上线 Hugging Face 和 GitHub。
官方给出的参数规模为 0.8B，定位是单一模型同时覆盖文档解析、关键信息抽取与 OCR 视觉问答三类任务。
在文档解析方向，该模型可识别正文、表格与数学公式，按阅读顺序输出结构化结果，原生支持 Markdown、LaTeX 格式。关键信息抽取面向发票、收据、卡证和业务表单，依据指定字段定位并输出 JSON 等结构化数据，可接入自动化处理流程。
OCR 视觉问答则允许围绕文档与图表内容进行跨区域检索、数值计算和语义推理，而非仅做文字转写。小米方面表示，这一组合覆盖了从底层文字感知到高层内容理解的链路。
#Xiaomi#Xiaomi-OCR-0#Qwen3.5-0.8B-Base#HunyuanOCR-1.5#TokenVL-2B#TextHawk2#Qwen3.5-2B#MiniCPM-V-4.5#Qwen3.5-4B#Super Xiaoai