Xiaomi open-sources Xiaomi-OCR-0, a 0.8B document model scoring 96.83 on OmniDocBench v1.6

Xiaomi has open-sourced Xiaomi-OCR-0, a 0.8-billion-parameter document understanding model that, according to Xiaomi, scores 96.83 on OmniDocBench v1.6, first among models under 1 billion parameters. Weights and inference code are available on Hugging Face and GitHub, the company said on October 30.

One model, three tasks

Xiaomi-OCR-0 handles three task types in a single model:

• Document parsing: recognizes body text, tables and math formulas and outputs structured results in reading order, with native Markdown and LaTeX support.
• Key information extraction: targets invoices, receipts, ID cards and business forms, locating requested fields and returning JSON for automated workflows.
• OCR visual question answering: goes beyond transcription to cross-region lookup, numerical calculation and reasoning over documents and charts.

Training

Xiaomi says the model is built on Qwen3.5-0.8B-Base and trained on about 170 million OCR and document samples. Labels were filtered through a consensus of several expert OCR models that cross-checked and voted on text, table and formula regions.

For difficult tables and formulas, the team used a predict, re-render and compare loop. Candidate outputs were rendered back into images and compared with the original at pixel and structure level to find misaligned rows or missing symbols. Synthetic hard cases added multilingual text, historical handwriting, screen moire and distorted lighting.

Training ran in three stages:

1. Text anchoring, to align visual regions with text tokens.
2. Continued OCR pretraining on tens of billions of tokens.
3. Mixed-task reinforcement learning with verifiable rewards, such as structural edit distance and table-matching scores.

Benchmarks (all figures Xiaomi's own)

Benchmark	Score
OmniDocBench v1.6	96.83
Real5-OmniDocBench	95.24
Wild-OmniDocBench	87.94
UniMER-Test (formulas)	97.84
PubTabNet (table structure)	92.27
Average of DocVQA, InfoVQA, ChartQA, OCRBench, TextVQA	83.2

On OmniDocBench v1.6, Xiaomi-OCR-0 ranks second overall, 0.04 points behind the leader. Real5-OmniDocBench simulates scanning, curled paper, screen photos, lighting changes and tilted shots.

On the five-benchmark VQA average, the 83.2 score is ahead of HunyuanOCR-1.5 at 76.5 and the 8B MiniCPM-V-4.5 at 82.6. It is behind Qwen3.5-4B at 84.9.

Deployment

Xiaomi says the technology already runs in the Super XiaoAi assistant on its phones, pulling pickup codes, queue numbers, bill details and meeting times from the screen. Alongside the weights, Xiaomi offers a browser demo, an MCP service and an Agent Skill toolkit for developers.

Context window and API pricing are not yet disclosed.

What this means

Xiaomi-OCR-0 puts parsing, field extraction and document question answering into one 0.8B model, so a single small checkpoint can cover work that is often split across separate tools. Its strongest reported results are on parsing: second overall on OmniDocBench v1.6 and first under 1B parameters. On the VQA average, it trails the 4B Qwen3.5 model, so the reported lead is in parsing rather than question answering.

The predict, re-render and compare loop and the verifiable rewards (structural edit distance, table-matching scores) tie training directly to structural accuracy of tables and formulas. The deployment in Super XiaoAi on Xiaomi phones, plus the MCP service and Agent Skill toolkit, gives developers working integration paths alongside the open weights.