Xiaomi open-sources 0.8B parameter OCR model Xiaomi-OCR-0 scoring 96.83 on OmniDocBench
Xiaomi announced on October 30 that it had open-sourced the lightweight document understanding model Xiaomi-OCR-0, with weights and inference code published on Hugging Face and GitHub.
Xiaomi announced on October 30 that it had open-sourced the lightweight document understanding model Xiaomi-OCR-0, with weights and inference code published on Hugging Face and GitHub. At 0.8B parameters, the single model covers document parsing, key information extraction and OCR visual question answering. It can recognize body text, tables and mathematical formulas, output structured results in reading order and natively support Markdown and LaTeX. Information extraction targets invoices, receipts, cards and business forms, locating fields per specification and emitting JSON so results can feed automated pipelines, while OCR visual question answering supports cross-region retrieval, numerical computation and semantic reasoning over documents and charts. The model is built on the Qwen3.5-0.8B-Base backbone and was trained on about 170 million OCR and document understanding samples, using a heterogeneous expert consensus scheme where multiple OCR experts cross-check and vote on text, tables and formulas to keep high-confidence annotations. For complex tables and formulas the team introduced a predict-rerender-visually-compare loop, re-rendering candidate parses into images and comparing them with the original regions to locate misaligned rows and columns or missing symbols before targeted correction, and synthesized hard samples adding multilingual, ancient handwriting, screen moire and distorted lighting data. Training ran in three stages covering Q-Mask text anchoring, continued OCR pretraining and mixed-task reinforcement learning. Xiaomi reports 96.83 on the combined OmniDocBench v1.6 ranking second overall just 0.04 behind the leader and first among sub-billion-parameter models, 97.84 on UniMER-Test, 92.27 on PubTabNet and 73 on the Chronicles-OCR ancient script set, level with the closed Seed 2.0 Pro at 72. The company says the technology is already used in the Super XiaoAi assistant on phones.