Xiaomi-OCR-0
Xiaomi open-sourced Xiaomi-OCR-0, a 0.8B document model that scores 96.83 on OmniDocBench v1.6, first among models under 1B parameters.
One 0.8B model handles document parsing, key information extraction and OCR visual question answering. It is built on Qwen3.5-0.8B-Base and trained on about 170 million samples in three stages, ending with reinforcement learning using verifiable rewards such as structural edit distance and table-matching scores. Xiaomi says it already runs in the Super XiaoAi assistant on its phones.
- Date
- Friday 30 October 2026
- Lab
- Xiaomi
- Kind
- open-weights
- Access
- open weights
Figures
| Measure | Value | Measured by |
|---|---|---|
| OmniDocBench v1.6 | 96.83 second overall, 0.04 behind the leader, first under 1B | company |
| Real5-OmniDocBench | 95.24 simulated scans, curled paper, screen photos | company |
| Wild-OmniDocBench | 87.94 | company |
| UniMER-Test | 97.84 formula benchmark | company |
| PubTabNet | 92.27 table structure | company |
| Average of DocVQA, InfoVQA, ChartQA, OCRBench, TextVQA | 83.2 HunyuanOCR-1.5 76.5, MiniCPM-V-4.5 8B 82.6, Qwen3.5-4B 84.9 | company |
A secondary source says the weights are Apache 2.0 and that the Real5 lead depends on the scoring code. All figures are Xiaomi's own.
Sources
- pandaily.com/xiaomi-ocr-0-open-source-0-8b-document-parsing-key-information-extraction-omn
- pasqualepillitteri.it/en/news/28641/xiaomi-ocr-0-open-source-ocr-model-documents
This record was checked against its sources on 31 October 2026. How we check