Xiaomi releases Xiaomi-OCR-0, an open-source 0.8B parameter OCR model

Summary

Xiaomi has open-sourced Xiaomi-OCR-0, a document understanding model built on the Qwen3.5-0.8B-Base infrastructure with 0.8 billion parameters. The model unifies document parsing, structured information extraction, and OCR visual quality control under one roof, offering local Markdown, LaTeX, and JSON outputs. Trained on approximately 170 million examples, the model’s weights, code, and MCP/Agent toolkit have been published on Hugging Face and GitHub.

Why it matters

A sub-1-billion parameter model that can compete with 2–8 billion parameter rivals in document benchmarks is genuinely effective in on-device and edge deployment scenarios where latency and memory constraints are decisive. This approach, which combines parsing, extraction, and quality control into a single open-weight model, lowers the integration barrier for developers who previously required a pipeline of specialized tools. The addition of MCP and Agent Skill interfaces indicates that Xiaomi is positioning this model not just as a standalone OCR API, but as a building block for agentic workflows.

Highlights

• 0.8 billion parameters, approximately 170 million training examples; a three-stage training process based on a predict-redraw-visual comparison loop for tables and formulas
• According to their own report, a score of 96.83 on OmniDocBench v1.6 (2nd place, 0.04 points behind the leader) and 95.24 on Real5-OmniDocBench (1st place among compared models)
• Document understanding average of 83.2, outperforming 1–7 billion parameter OCR specialists; however, it falls below Qwen3.5-4B (84.9) and MiniCPM-V-4.5 (82.6). A score of 73 in ancient writing systems remains modest.

Critical take

All benchmark data relies on Xiaomi’s self-reported claims; there is no independent reproduction. The '1st place' claims in the headline are limited to the narrow comparison sets chosen by Xiaomi. The 0.04-point difference from the OmniDocBench leader is statistically insignificant. The score of 73 on Chronicles-OCR reveals that, despite the claim of 170 million training examples, long-tail, low-resource writing systems remain a weak point.