On October 30, Xiaomi formally promoted its open-source document model Xiaomi-OCR-0 through its official accounts. The model has 0.8B parameters, is built on Qwen3.5-0.8B and is released under the Apache-2.0 license. The weights and technical report actually went public at the end of September; this time Xiaomi added an MCP server and an Agent Skill, making it a proper developer-facing launch.
Xiaomi says the model ranks first among models under 1 billion parameters in the OmniDocBench v1.6 document parsing test. That claim matches the table in the README: Xiaomi-OCR-0 scores 96.83, with only the 1.2B TeleOCR slightly ahead at 96.87.
One model, three tasks
Xiaomi-OCR-0 handles document parsing, key information extraction and text-centric visual question answering in a single model. According to the technical report, the training corpus holds roughly 170 million samples produced by an automated data engine that combines expert consensus, rendering-based verification and targeted synthesis. Training starts from Qwen3.5-0.8B and runs through Q-Mask text anchoring, continued pre-training and mixed-task reinforcement learning.
Xiaomi's other published results include 97.84 on UniMER-Test formula recognition, a TEDS of 92.27 on PubTabNet table recognition, and an average of 83.2 across five visual question answering benchmarks, compared with 80.9 for Qwen3.5-2B. All of these are the developer's own measurements.
For deployment, the model runs on SGLang or vLLM, and there is an experimental Ollama package. The README includes a prompt that lets an agent read SKILL.md and install the tooling itself. The page-by-page layout parsing mode requires extra installs of PaddlePaddle, PaddleX and PP-DocLayoutV3.
What separates 95.24 from 92.73
More interesting than the “No. 1” claim is the Real5-OmniDocBench column. This dataset takes the 1,355 document pages from OmniDocBench v1.5 and recaptures each under five conditions: scanning, warping, screen re-photographing, lighting changes and tilt. That yields 6,775 images, closer to real-world photographed documents than clean digital files.
In the README table, Xiaomi scores 95.24, PaddleOCR-VL-1.6 gets 93.19 and OvisOCR2 gets 92.29, putting Xiaomi first. The public leaderboard on Hugging Face shows a different order:
| Rank | Model | Parameters | Leaderboard score |
|---|---|---|---|
| 1 | PaddleOCR-VL-1.6 | 0.9B | 93.19 |
| 2 | Xiaomi-OCR-0 | 0.8B | 92.73 |
| 3 | OvisOCR2 | 0.9B | 92.29 |
| 5 | GLM-5.3-Flash | 321B | 90.76 |
| 8 | Kimi-K2.6 | 1.1T | 89.76 |
| 9 | Gemini-3 Pro | Not disclosed | 89.24 |
The leaderboard discussion thread records how it unfolded. The party that submitted the evaluation said Real5 has to be scored with the v1.5 branch of OmniDocBench. A Xiaomi researcher replied that the company's reference figures were produced with v1.6 code: 95.24 for the two-stage pipeline and 93.45 end to end, and then handed over the raw predictions from the two-stage run. The other side rescored them with v1.5 and got an overall 92.7863, essentially matching the leaderboard figure. That result was later merged into the leaderboard.
Both sides agree the evaluation code version is the main factor, and Xiaomi also mentioned that vLLM inference parameters may play a role. The thread does not break down how much of the gap between the two versions comes from the code alone.
The leaderboard itself is still in flux
The problem with the README table is mixed methodology: Xiaomi's cell is a score from the v1.6 code, while the 93.19 and 92.29 figures for its rivals match the leaderboard. Compared in the same column, the claim to first place does not hold up.
There is some context to this dispute as well. The Real5 dataset belongs to the PaddlePaddle organization, the top-ranked model is its own PaddleOCR-VL-1.6, and PP-DocLayoutV3, the layout detector used in Xiaomi's two-stage approach, comes from the same ecosystem. Nothing in the thread suggests the scoring is biased, and the rescoring was done in the open, but we found no independent third-party reproduction.
In the discussion, a participant from the dataset side said Real5 will soon be upgraded to align with the latest OmniDocBench and publish unified new metrics. Until then, teams choosing a model gain little from drawing conclusions off a 0.4-point gap. All of these models can be downloaded for free, and running them on your own scans and phone photos will tell you more than the leaderboard.
Sources: Xiaomi Research GitHub repository and technical report; CocoLoop; Real5-OmniDocBench public leaderboard and Hugging Face discussion. Except for the leaderboard, all scores are developer-reported.