Xiaomi-OCR-0: A Unified 0.8B Vision-Language Model for Document Parsing and OCR

Xiaomi-OCR-0 is a unified 0.8B vision-language model designed for document parsing and OCR-centric understanding. Built upon Qwen3.5-0.8B, the model has been trained on an approximately 170M-sample OCR-centric corpus utilizing Q-Mask text anchoring, continued pretraining (CPT), and mixed-task reinforcement learning (Mix-RL).

Capabilities and Performance

The model handles multiple document tasks and reports various benchmark results:

• Document Parsing: Produces structured Markdown from document pages, converting OTSL outputs to HTML tables and formulas to LaTeX.
• Key Information Extraction (KIE): Extracts requested fields into JSON format.
• Visual Question Answering (VQA): Answers queries regarding text and content within document images.

Installation and Local Runtimes

Running Xiaomi-OCR-0 locally requires Python 3.10 or newer along with an inference runtime such as SGLang, vLLM, or Ollama.

• Browser Demo: A local web interface is available for testing whole-page and region document parsing, KIE, VQA, and PDF parsing without installing full agent skills.
• Region Parsing Dependencies: Region mode requires additional layout dependencies including PaddlePaddle, PaddleX, and PP-DocLayoutV3 weights.
• Token Limits: The default output limit is 4096 tokens for document parsing and KIE, while VQA uses a fixed 1024-token limit.

Community conversions, such as GGUF and Ninfer packages, are available via contributors, though users should note that original benchmark scores do not automatically apply to these alternative weight formats.