VarwoXiaomi-OCR-0: Model, Demo and Document Parsing Guide
Find Xiaomi-OCR-0 weights and source code, choose an OCR output format, and evaluate document parsing without mistaking model benchmarks for your own results.
Official model card · Source repository · Technical report
Xiaomi-OCR-0 is a 0.8B vision-language model for document parsing and OCR-related understanding. It is useful to evaluate when you need text, document structure or specific fields from an image. Start with the official model card for weights and supported usage; the research report explains how it was trained and evaluated.
Sources checked November 3, 2026. This is a documentation guide. Varwo has not downloaded the weights, run inference, tested hardware requirements or reproduced the benchmarks.
Choose the output before the runtime
The model card documents different tasks:
| Input and task | Documented output |
|---|---|
| Whole document page | Markdown in reading order, with OTSL tables and LaTeX formulas |
| Text crop | Plain text |
| Table crop | OTSL table representation |
| Formula crop | LaTeX |
| Named fields with a supplied schema | A JSON object |
OTSL output needs a compatible downstream reader; do not assume it is already a spreadsheet or a conventional Markdown table. Decide how your application will consume the result before choosing a sample prompt from the card.
Where to try or deploy it
Use the official repository to reach the project’s current examples and demo links. The older SeerRay-Lab repository URL redirected to xiaomi-research/Xiaomi-OCR-0 when checked. A demo link is a route to investigate, not a promise of an available GPU session or free unlimited processing.
For local evaluation, follow the card’s current Transformers instructions and distinguish the original weights from community conversions. The card labels the model Apache-2.0; check the chosen conversion and its dependencies separately. Parameter count alone does not establish how much memory your chosen runtime, image size and output length will require.
A small evaluation you can repeat
Prepare a few non-sensitive examples with manually checked answers: a clean page, a skewed photo, a dense table and a page with formulas. Record missing lines, reading-order errors, incorrect cells and malformed output separately. Keep image preprocessing and output settings consistent when comparing models.
The technical report reports 96.83 on OmniDocBench v1.6 and 95.24 on Real5-OmniDocBench. These are the authors’ benchmark results, not percentages of your documents guaranteed correct. Verify consequential fields against the original image before using them downstream. For a broader PDF workflow, see PDFCraft.