What Xiaomi has made public
Xiaomi has announced the opening of the Xiaomi-OCR-0 model, a 0.8-billion-parameter system designed to combine three distinct functions: structured document reading, key-field extraction and so-called visual OCR, that is, the ability to answer questions about texts, tables and charts. According to IT之家, the model weights and inference code were published simultaneously on Hugging Face and GitHub, with the addition of an interactive browser demo, an MCP service and an Agent Skill toolkit for integrating the model into existing workflows.
The choice of a lightweight format is not marginal: the reduced size makes the model easier to run and adapt than the large multimodal systems cited in the comparison. The same source specifies that the model derives from the Qwen3.5-0.8B-Base base and that training used approximately 170 million samples across OCR and document understanding.
The three declared functions and the training pipeline
On the parsing front, the model recognizes text, tables and mathematical formulas and returns the reading order in formats such as Markdown and LaTeX. For information extraction, by contrast, it identifies specific fields in invoices, receipts, cards and corporate forms, producing structured outputs such as JSON. The third function, visual OCR, makes it possible to query a document's content beyond simple transcription, with requests that may require searching across different areas, calculations or semantic inferences.
The source material also describes some technical steps in training. To reduce labeling errors, Xiaomi used a consensus mechanism among OCR experts that compares and votes on high-confidence annotations. On complex tables and formulas, a circular check was introduced: the hypothesized result is rendered again as an image and compared with the original region, so as to identify shifted rows or columns and missing symbols. Difficult examples were also generated to cover scans, paper curvature, reflections, tilted photographs, handwriting and historical materials.
The benchmark numbers and the limits of interpretation
According to results released by Xiaomi, Xiaomi-OCR-0 scores 95.24 points on the Real5-OmniDocBench benchmark, which simulates five real-world acquisition conditions, and 87.94 on Wild-OmniDocBench. On the general OmniDocBench v1.6 benchmark it reaches 96.83 points, second place overall, 0.04 points from the top and first place among models under one billion parameters. In specific tests it reports 97.84 on UniMER-Test for formulas and 92.27 on PubTabNet for table structure.
On document understanding, the average across the five benchmarks DocVQA, InfoVQA, ChartQA, OCRBench and TextVQA is given as 83.2 points, a value Xiaomi compares with HunyuanOCR-1.5, TokenVL-2B and TextHawk2, as well as with general-purpose multimodal models. It should be emphasized, however, that these are data provided by the company and not independent verifications reported in the material: comparability depends on the test configurations, which the source does not detail. It therefore remains to be confirmed both how it performs on real documents different from the sample and its actual behavior in third-party integrations.
Where it is already used and what is missing
The source indicates an already active use in the Super Xiaomi voice and visual assistant: with a short press of the AI button, a three-finger swipe or a voice command, the system extracts pickup codes, queue numbers and temporary information from the screen. The same mechanism is described for reading transactions and merchant data, useful for automatic expense management, and for filling in fields from invoices and identity documents.
The material does not, however, indicate the license terms, hardware requirements for running it, integration costs or any restrictions on commercial use. These elements are decisive for assessing the actual openness of the project, which at the moment can be judged only on the basis of what has been announced and the published references.
