Xiaomi open-sources 800000000 parameters OCR model supporting complex table parsing and document question answering

AT A GLANCE

Xiaomi has open-sourced 800000000 parameters of Xiaomi-OCR-0, which can parse documents, extract fields, and perform visual question answering.

Article

Xiaomi open-sources 800000000 parameters OCR model that can read complex tables, extract information, and answer questions

Beating AI breaking news: Xiaomi has open-sourced Xiaomi-OCR-0, with only 800000000 parameters, capable of completing types of tasks: converting text, complex tables, and mathematical formulas in documents into editable content; extracting specified fields from invoices, receipts, and identity documents and outputting JSON; and answering questions based on information in documents and charts, rather than merely transcribing text from images.

The model was further trained based on Alibaba's Qwen3.5-0.8B, using approximately 170000000 OCR and document-understanding samples. To improve data quality, Xiaomi used multiple OCR models to cross-check recognition results. For complex tables and formulas, it also re-rendered the recognition results as images and compared them with the originals to identify issues such as misalignment and omitted characters. Training was also conducted in stages: first improving text recognition and layout analysis, then adding tasks such as information extraction and visual question answering.

In OCR visual question-answering tests released by Xiaomi,Xiaomi-OCR-0 achieved an average score of 83.2, surpassing 2.5 with Qwen3.5-2B（80.9), and approaching 4B model's 84.9. In Real5 document-parsing tests involving interference such as simulated scanning, paper bending, and screen recaptures, it achieved 95.24 points when paired with an external layout detector.

The model weights, inference code,MCP service, and Agent Skill have all been open-sourced. Developers can integrate it with AI Agent to read screenshots,PDF and receipts, extracting data that can be directly processed by other programs.

Original link https://m.theblockbeats.info/flash/374762

Key points

Xiaomi-OCR-0 has 800000000 parameters and can convert document text, complex tables, and mathematical formulas into editable content.
The model can extract specified fields from invoices, receipts, and identity documents, and output JSON。
The model can answer questions based on information in documents and charts.
The model was further trained based on Alibaba's Qwen3.5-0.8B using approximately 170000000 OCR and document-understanding samples.
In OCR visual question-answering tests released by Xiaomi,Xiaomi-OCR-0 achieved an average score of 83.2，Qwen3.5-2B of 80.9，4B model was 84.9。
In Real5 document-parsing tests, the model achieved 95.24 points when paired with an external layout detector.
The model weights, inference code,MCP service, and Agent Skill have all been open-sourced.

AI-assisted interpretation

The following is analysis, separate from reported facts. Verify important claims independently.

This is a model for processing images and documents. It can not only recognize text, but also understand tables, formulas, and layouts; organize specified information from receipts; and answer questions based on document content. Outputting JSON means that the extracted results can be more conveniently handed off to other programs for further processing.

Why it matters to readers

If its real-world performance matches the report, developers can integrate it with AI Agent to read screenshots,PDF and receipts, reducing manual data entry and document organization.

Beginners can think of it as an open-source “document-reading assistant”; however, actual use may still involve model deployment, inference code, and a layout detector.

Risks and unknowns

The test results were released by Xiaomi, and the evidence contains no independent verification results.
The evidence does not provide addresses for the model repository, open-source license, or technical report.
Real5 The test results used an external layout detector, so this article alone cannot determine the model's performance when run independently.
The evidence does not specify the model's accuracy, operating costs, or hardware requirements in actual business use.

Xiaomi open-sources 800000000 parameters OCR model supporting complex table parsing and document question answering

Article

AI-assisted interpretation

Why it matters to readers

Risks and unknowns

Related Developments

Related concepts

Keep following

Security Center