Built on Qwen3.5-0.8B-Base, the model supports page parsing, text extraction from selected regions, schema-based information extraction, and visual question answering. Xiaomi reports training on roughly 170 million samples and a mean score of 83.2 across five OCR-VQA benchmarks.

Weights and GGUF quantizations are available, with Transformers support documented. The roughly 1.8 GB Ollama package is an experimental community build. The reported results are company-supplied benchmark scores.

Why it matters

A compact model can be tested for OCR and structured extraction on documents using local or established inference setups.

GUIDES

Something to try

As a hypothetical exercise, test a non-sensitive scanned document and check whether tables and reading order are preserved.