Xiaomi opens a 0.8-billion-parameter document model
Prepared by Sintelligent with the help of SI. Published in SI: . Updated: .
A small local model can turn document pages into data for other programs.
OCR-0 combines text recognition, field extraction and document question answering, producing Markdown, LaTeX and JSON. It is built on Qwen3.5-0.8B and trained on around 170 million examples; code and weights are available under Apache 2.0. The package includes tools for local use and agent integration.
Sources 2