Xiaomi open-sources the 0.8B-parameter OCR model Xiaomi-OCR-0, supporting Markdown and JSON output

Xiaomi Open-Sources 0.8B-Parameter OCR Small Model: Bridging Document Parsing and Visual Question Answering, Supports Markdown and JSON Output

2026-10-30 21:15 IT Home - Wenzhou

Thanks to IT Home users Pingdi Guoguo and Youji Xuehu for the tip-off!

IT Home, October 30 news: Xiaomi officially announced today the open-sourcing of its lightweight document understanding model Xiaomi-OCR-0, with the model weights and inference code now available on Hugging Face and GitHub.

The official parameter count is 0.8B, positioned as a single model covering three task types at once: document parsing, key information extraction, and OCR visual question answering.

In the document parsing direction, the model can recognize body text, tables, and mathematical formulas, outputting structured results in reading order, with native support for Markdown and LaTeX formats. Key information extraction targets invoices, receipts, cards and certificates, and business forms, locating and outputting structured data such as JSON according to specified fields, and can be integrated into automated processing workflows.

OCR visual question answering allows cross-region retrieval, numerical computation, and semantic reasoning around document and chart content, rather than merely performing text transcription. Xiaomi said this combination covers the chain from low-level text perception to high-level content understanding.

According to Xiaomi's official disclosure, Xiaomi-OCR-0 is built on the Qwen3.5-0.8B-Base foundation, with cumulative training on approximately 170 million OCR and document understanding samples. On the data side, it adopts a heterogeneous expert consensus mechanism, in which different regions such as body text, tables, and formulas are cross-validated by multiple OCR expert models and high-confidence annotations are selected by voting.

For complex tables and formulas, the team introduced a closed-loop validation of "predict - re-render - visual comparison": candidate parsing results are re-rendered as images and compared with the original document regions at the pixel and structure level, then targeted corrections are made after locating row-column misalignment or missing symbols. In addition, difficult samples are synthesized according to layout complexity, font morphology, and physical degradation factors, supplementing long-tail data such as multilingual text, ancient handwritten scripts, screen moiré patterns, and distorted lighting.

Training is divided into three stages: Q-Mask text anchoring strengthens the spatial correspondence between local visual regions and text tokens; OCR continual pre-training lays the foundation with multi-task mixed data at the tens-of-billions-of-tokens level; and mixed-task reinforcement learning uses structural edit distance, table matching degree, and question-answering accuracy as verifiable reward signals on difficult samples.

Xiaomi's published evaluation results show that on Real5-OmniDocBench, which simulates five real capture methods - scanning, paper bending, screen re-photography, lighting changes, and tilted shooting - Xiaomi-OCR-0 scores 95.24 points, ranking first among the models it was compared with. On Wild-OmniDocBench it scores 87.94 points.

On the comprehensive benchmark OmniDocBench v1.6, the model scores 96.83 points, ranking second on the overall leaderboard, 0.04 points behind the top spot, and first among models with fewer than 1B parameters. In specialized tests, it scores 97.84 points on the mathematical formula recognition benchmark UniMER-Test and 92.27 points on the table structure recognition benchmark PubTabNet.

On the Chronicles-OCR (Mature Scripts) evaluation set for ancient texts and calligraphy, Xiaomi-OCR-0 scores 73 points, which Xiaomi says is on par with the closed-source model Seed 2.0 Pro's 72 points.

In document understanding, Xiaomi averaged scores across five benchmarks - DocVQA, InfoVQA, ChartQA, OCRBench, and TextVQA - with Xiaomi-OCR-0 at 83.2 points.

Official comparison data shows this result is higher than HunyuanOCR-1.5 (1B, 76.5 points), TokenVL-2B (2B, 78.1 points), and TextHawk2 (7B, 78.5 points).

In the same comparison group, the general multimodal model Qwen3.5-2B scores 80.9 points, the 8B-scale MiniCPM-V-4.5 scores 82.6 points, and Qwen3.5-4B scores 84.9 points.

In bill and card scenarios, screen recognition can extract transaction amounts, merchant names, payment times, and spending categories for smart bookkeeping; paper invoices, ID cards, bank cards, and other materials can have fields extracted by taking photos. In scheduling and travel scenarios, the system can recognize meeting notices in group chats or the time and location on posters and generate schedule cards.

IT Home notes that in addition to releasing the 0.8B model weights and inference code, Xiaomi also provides a browser-based interactive demo, an MCP service, and an Agent Skill toolkit, which developers can integrate into their own business workflows or agent systems.

Reference materials: