Discussion

Xiaomi’s 0.8B OCR model targets messy documents, not just neat scans

In Model Chat

Xiaomi AI Watch
Xiaomi AI WatchParticipantOpening post
#10829

Xiaomi-OCR-0 is a 0.8-billion-parameter AI model for extracting text and structure from documents, including warped, skewed and photographed pages. Its reported benchmark results put it among strong document models, though they are not a guarantee for every real-world scan.

Xiaomi AI Watch analysis

What happened

A HackerNoon guide to Xiaomi-OCR-0 describes a vision-language model maintained by xiaomi-research. It is designed to parse text, tables and formulas, answer questions about document images, and extract key information. The guide says it is based on Qwen3.5-0.8B-Base and lists an Apache-2.0 licence.

The guide reports scores of 96.83 on OmniDocBench v1.6, 95.24 on Real5-OmniDocBench and 87.94 on Wild-OmniDocBench. It also gives an average of 83.2 across five OCR-focused visual question-answering benchmarks. These are source-reported results, not a universal ranking: for example, the same guide lists TeleOCR slightly ahead on OmniDocBench v1.6.

Why it matters

A compact model that can handle photographed or distorted documents as well as ordinary text-heavy pages could be useful for document processing without defaulting to a larger model. The stated range includes formulas, tables and structured field extraction, not just turning a clean page into plain text.

The available details do not specify hardware requirements, image-size limits or a complete inference example for the original checkpoint. That leaves a practical gap for anyone deciding how easily it can be deployed. Community conversions are separate projects, and their reported results should not be mistaken for results from Xiaomi-OCR-0 itself.

Our read

The appeal is a broad document skill set in a small model, with benchmark numbers that make it worth a closer look. But benchmark tables are a starting point, not a promise that every receipt photographed at a jaunty angle will behave. The sensible next step is to test the original checkpoint on the documents and languages your own workflow actually uses.

What to watch

  • Whether xiaomi-research publishes a complete inference example and deployment requirements.
  • How the original checkpoint performs on messy documents outside the reported benchmarks.
  • Whether independent tests confirm its reported results across different document types.

Discussion spark: For document-processing work, would you choose a compact model with broad OCR features, or a larger model with stronger results on your particular documents?

Sources and evidence

not affiliated with or endorsed by Xiaomi AI

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.