Article · AI translation
Thanks to IT Home readers 平底锅锅 and 有鲫雪狐 for the tips!
IT Home reported on October 30 that Xiaomi officially announced today the open sourcing of its lightweight document understanding model Xiaomi-OCR-0. The model weights and inference code are now available on Hugging Face and GitHub.
According to the official specifications, the model has 0.8B parameters and is designed to cover three types of tasks with a single model: document parsing, key information extraction, and OCR visual question answering.
For document parsing, the model can recognize body text, tables, and mathematical formulas, output structured results in reading order, and natively support Markdown and LaTeX formats. Key information extraction targets invoices, receipts, cards, certificates, and business forms; it locates specified fields and outputs structured data such as JSON, which can be integrated into automated workflows.
OCR visual question answering allows cross-region retrieval, numerical calculation, and semantic reasoning about document and chart content, rather than only text transcription. Xiaomi said this combination covers the process from basic text perception to higher-level content understanding.
According to Xiaomi, Xiaomi-OCR-0 is built on Qwen3.5-0.8B-Base and was trained on a total of about 170 million OCR and document understanding samples. For data processing, it uses a heterogeneous expert consensus mechanism: multiple OCR expert models cross-check different regions, including body text, tables, and formulas, and vote to select high-confidence annotations.
For complex tables and formulas, the team introduced a closed-loop verification process of “prediction—re-rendering—visual comparison”: candidate parsing results are rendered back into images and compared with the original document regions at the pixel and structural levels, allowing misplaced rows or columns and missing symbols to be identified and corrected. The team also synthesized difficult samples based on layout complexity, font characteristics, and physical degradation factors, supplementing long-tail data such as multilingual text, handwritten ancient texts, screen moiré patterns, and distorted lighting.
Training consists of three stages: Q-Mask text anchoring strengthens the spatial correspondence between local visual regions and text tokens; continued OCR pretraining provides a foundation using multitask mixed data containing tens of billions of tokens; and mixed-task reinforcement learning uses structural edit distance, table matching, and question-answering accuracy as verifiable reward signals on difficult samples.
Xiaomi's published evaluation results show that Xiaomi-OCR-0 scored 95.24 points on Real5-OmniDocBench, which simulates five real-world capture methods: scanning, paper bending, photographing a screen, lighting changes, and tilted photography. It ranked first among the models Xiaomi compared. It scored 87.94 points on Wild-OmniDocBench.
On the comprehensive OmniDocBench v1.6 benchmark, the model scored 96.83 points, ranking second overall, 0.04 points behind the leader, and first among models with fewer than 1B parameters. In specialized tests, it scored 97.84 points on the mathematical formula recognition benchmark UniMER-Test and 92.27 points on the table structure recognition benchmark PubTabNet.
On the ancient texts and calligraphy evaluation set Chronicles-OCR (Mature Scripts), Xiaomi-OCR-0 scored 73 points. Xiaomi said this was on par with the closed-source model Seed 2.0 Pro, which scored 72 points.
For document understanding, Xiaomi averaged scores across five benchmarks—DocVQA, InfoVQA, ChartQA, OCRBench, and TextVQA—and Xiaomi-OCR-0 scored 83.2 points.
According to official comparison data, this score exceeds those of HunyuanOCR-1.5 (1B, 76.5 points), TokenVL-2B (2B, 78.1 points), and TextHawk2 (7B, 78.5 points).
In the same comparison, the general-purpose multimodal model Qwen3.5-2B scored 80.9 points, the 8B MiniCPM-V-4.5 scored 82.6 points, and Qwen3.5-4B scored 84.9 points.
Regarding practical use, Xiaomi said the technology has been incorporated into the phone-based “Super XiaoAI.” By briefly pressing the AI button, swiping up with three fingers, or giving a voice command, users can extract time-sensitive information from the screen, such as food pickup codes, package collection codes, and queue numbers, and place it in a prominent position.
For bills and cards, the screen recognition feature can extract transaction amounts, merchant names, payment times, and spending categories for smart expense tracking; fields can also be extracted from photographed paper invoices, ID cards, and bank cards. For scheduling and travel, the system can identify meeting notices in group chats or times and locations on posters and generate calendar cards.
IT Home noted that, in addition to releasing the 0.8B model weights and inference code, Xiaomi provides an interactive browser demo, an MCP service, and an Agent Skill toolkit, allowing developers to integrate the model into their own workflows or agent systems.
References:
Hugging Face model weights: https://huggingface.co/xiaomi-research/Xiaomi-OCR-0
GitHub project code: https://github.com/xiaomi-research/Xiaomi-OCR-0
Agent one-step trial command: Read and execute https://raw.githubusercontent.com/xiaomi-research/Xiaomi-OCR-0/main/SKILL.md