AI knows how to read your PDFs. Xiaomi now wants it to understand them.

With Xiaomi-OCR-0, the Chinese manufacturer offers an open 800-million-parameter AI model capable of reading, structuring, and querying complex documents, with performance rivaling that of much larger models.

Recognizing the words in a document doesn't necessarily mean understanding its contents. An invoice contains amounts, addresses, and tax information. A scientific report combines paragraphs, equations, and tables. A photographed document adds challenges involving perspective, lighting, and layout. To process this information correctly, artificial intelligence needs to do more than transcribe the characters it can recognize. That's exactly the problem Xiaomi is tackling with Xiaomi-OCR-0, an 800-million-parameter model specialized in document understanding.

The Chinese technology company claims performance comparable to much larger models, reporting a score of 95.24 on Real5-OmniDocBench and an average of 83.2 across five visual question-answering benchmarks. The release of the model weights, inference code, and integration tools also allows developers to experiment with the technology in their own environments. Although Xiaomi highlighted the project again in an October 30 announcement, its technical report and initial model weights had already been published in late September 2026. Reading a Document Also Means Understanding Its Structure

Optical character recognition, commonly known as OCR, converts text contained in images or scanned documents into machine-readable characters. The technology has existed for decades, but today's requirements go far beyond transcription. Businesses want to locate specific information across thousands of documents, extract structured data, and enable AI assistants to work with archives without requiring extensive manual processing.

Xiaomi-OCR-0 combines three major capabilities within a single model. The first is document parsing: it can recognize different elements on a page, preserve their reading order, and produce a structured representation. The second is key information extraction, such as identifying an invoice's total amount, date, and supplier. The third is visual question answering, allowing users to ask questions about a document and receive answers based on its visual content.

That third capability represents an important departure from traditional OCR. Instead of simply reproducing detected text, the model must connect a question to the relevant information, interpret the arrangement of elements, and generate the appropriate answer. A user could ask which column of a table contains a particular value or retrieve specific information from a complex form.

Outputs can take several forms depending on the task: Markdown for reconstructed documents, LaTeX for mathematical formulas, structured representations for tables, and JSON for extracting specific fields. This versatility is especially relevant to business applications that need to transform heterogeneous documents into usable data. 800 Million Parameters, 170 Million Training Examples

To build Xiaomi-OCR-0, researchers started with Qwen3.5-0.8B-Base, a compact multimodal model. They specialized it using a dataset of approximately 170 million samples focused on document recognition and understanding. This scale is one of the project's defining characteristics: Xiaomi is attempting to achieve strong performance from a relatively lightweight model through extensive task-specific training.

The process involves several complementary stages. The first, called Q-Mask, aims to improve the association between text elements and their locations within an image. This is followed by additional document-focused pretraining and a reinforcement learning stage combining multiple tasks with verifiable rewards.

The challenge isn't simply to increase the number of examples. Training data must also be reliable enough to teach the model how to recognize text, tables, unusual characters, and mathematical expressions accurately. Xiaomi therefore developed a data preparation system involving several specialized models.

These systems generate different interpretations of the same document, which are then compared. When their outputs disagree, annotations can be visually reconstructed and checked against the original document to identify errors. Synthetic examples supplement the dataset, particularly to strengthen underrepresented categories and address recurring mistakes observed during training.

The approach illustrates an important trend in compact AI development: performance depends as much on data quality and training methodology as it does on parameter count. Strong Results, but the Number One Ranking Needs Context

Xiaomi reports a score of 96.83 on OmniDocBench v1.6, a benchmark evaluating document parsing across text, formulas, tables, and complex page layouts. In the comparisons published by the research team, the model outperforms several similarly sized competitors, including OvisOCR2 and PaddleOCR-VL-1.6.

Its reported score on Real5-OmniDocBench reaches 95.24. This benchmark is particularly interesting because it evaluates documents captured under less-than-ideal conditions, including scanning artifacts, distortions, photographs of screens, inconsistent lighting, and tilted pages. It aims to reproduce some of the challenges encountered when documents come from smartphones, scanners, or imperfect photographs.

An important discrepancy appears in the public rankings, however. The Real5-OmniDocBench dataset leaderboard also displays a score of 92.73 for Xiaomi-OCR-0, different from the 95.24 reported in the technical documentation. Without an explicitly established connection between the evaluation configurations and submissions, those figures shouldn't be treated as interchangeable. Xiaomi's first-place claim therefore needs to be tied to the specific testing protocol and comparison table it uses.

Another important detail is that the reference document-parsing results rely on a two-stage process. An additional system, PP-DocLayoutV3, can first identify the different regions of a page before Xiaomi-OCR-0 processes them. The advertised 800 million parameters refer to the main model, not necessarily every component involved