Skip to content
Back to search

xiaomi-ocr-setup

SkillWorks with Claude CodeGitHub

Install and configure the Xiaomi-OCR-0 Skill and local MCP service for document OCR, KIE, and VQA.

Safety scan

A static text scan, not a guarantee. Select the badge to see what was scanned and every finding.

Checked today
Badge for your README
findsafeskills safety rating

Markdown

[![findsafeskills safety rating](/api/listings/xiaomi-research--xiaomi-ocr-0/badge.svg)](/listings/xiaomi-research--xiaomi-ocr-0)

HTML

<a href="/listings/xiaomi-research--xiaomi-ocr-0"><img src="/api/listings/xiaomi-research--xiaomi-ocr-0/badge.svg" alt="findsafeskills safety rating" /></a>

Install

i

Install instructions from the README

As written by the repo's author. findsafeskills has not verified these steps.

Choose one of two setup paths:

- [Try it with an agent](#try-it-with-an-agent): let your agent install the local model and configure the MCP service. - [Manual installation](#manual-installation): get the code, choose an inference runtime, and run the browser demo yourself.

Try it with an agent

> ⚠️ **Notice:** With your authorization, this Skill may create a Python environment, install dependencies, download model weights, install and configure the selected inference runtime (SGLang, vLLM, or Ollama), and register an MCP server. Model and layout downloads can be large. If you do not agree to this setup flow, do not execute the command below.

Copy and send this to your agent:

Read and execute https://raw.githubusercontent.com/SeerRay-Lab/Xiaomi-OCR-0/main/SKILL.md

The model will run on your machine. For a complete setup, the agent also installs the region-mode dependencies (PaddlePaddle, PaddleX and PP-DocLayoutV3) under your setup authorization and verifies region parsing. You can request page-only setup. If your platform does not support the layout runtime, the agent will explain the limitation.

Manual installation

Python 3.10 or newer is required. Get the repository first, then choose **one** runtime: [SGLang / vLLM](#sglang--vllm) or [Ollama](#ollama-experimental). To use region parsing, also install the [optional layout dependencies](#optional-region-mode-dependencies).

For manual MCP setup, see [INSTALL.md](INSTALL.md); Ollama-specific setup is covered in [OLLAMA.md](OLLAMA.md).

Get the repository

git clone https://github.com/SeerRay-Lab/Xiaomi-OCR-0.git
cd Xiaomi-OCR-0

SGLang / vLLM

You can run the browser demo without installing the Agent Skill or MCP server. The demo sends requests to a local SGLang or vLLM inference server at `http://127.0.0.1:8000/v1`.

1. Install an inference runtime

If a compatible local server is already running at the address above, skip to step 3. Otherwise, choose **one** runtime and follow its installation guide for your operating system, GPU, and driver. These runtimes have hardware-specific requirements; the linked guides list supported platforms and installation options:

- [SGLang installation guide](https://docs.sglang.ai/get_started/install.html) · [supported models](https://docs.sglang.io/docs/supported-models/multimodal_language_models.md) - [vLLM GPU installation guide](https://docs.vllm.ai/en/latest/getting_started/installation/gpu/) · [Qwen3.5 model support](https://docs.vllm.ai/en/latest/api/vllm/model_executor/models/qwen3_5/)

Use a release that supports the checkpoint architecture `Qwen3_5ForConditionalGeneration`. For macOS and non-NVIDIA hardware, check the runtime's platform-specific support before installing; the default commands below are for a compatible GPU installation.

2. Start Xiaomi-OCR-0

Run **one** of the following commands in a terminal. On first launch, the runtime downloads the checkpoint from Hugging Face if it is not already cached. Keep this terminal open while using the demo.

SGLang:

python -m sglang.launch_server --model-path SeerRay-Lab/Xiaomi-OCR-0 \
  --host 127.0.0.1 --port 8000 --context-length 16384

vLLM:

vllm serve SeerRay-Lab/Xiaomi-OCR-0 \
  --host 127.0.0.1 --port 8000 --max-model-len 16384

Wait for the server to finish loading the model before continuing.

3. Start the browser demo

Open a second terminal in the repository directory and run:

python3 -m venv .venv
source .venv/bin/activate
python -m pip install -r requirements.txt
python demo/server.py --port 8787

What was scanned

The files the safety scan read, as written by the repository's author and shown as plain text. Lines with a finding are marked.

SKILL.md · Manifest · 64 lines
  1. ---
  2. name: xiaomi-ocr-setup
  3. description: Install and configure the Xiaomi-OCR-0 Skill and local MCP service for document OCR, KIE, and VQA.
  4. ---
  5. # Xiaomi-OCR-0 Skill
  6. The only user-facing installation prompt is:
  7. ```text
  8. Read and execute https://raw.githubusercontent.com/SeerRay-Lab/Xiaomi-OCR-0/main/SKILL.md
  9. ```
  10. ## Install this Skill into the current agent
  11. If this file was read from the Xiaomi-OCR-0 repository and the user asked to install it,
  12. install only the `xiaomi-ocr` Skill globally for the current agent:
  13. ```bash
  14. npx skills add SeerRay-Lab/Xiaomi-OCR-0 --skill xiaomi-ocr --full-depth --global --agent <current-agent> --yes
  15. ```
  16. Replace `<current-agent>` with the current agent's Skills CLI identifier; do not install to every agent or modify a
  17. project-local skill directory. Then inspect the current agent's MCP configuration and local
  18. machine, and follow [backend selection in the runtime Skill](skills/xiaomi-ocr/SKILL.md#choose-the-inference-backend).
  19. Choose the service from the existing setup, hardware, and accuracy requirements; the user
  20. does not need to name a backend or model in the installation prompt. Apply the selected
  21. backend's environment variables before running the installer.
  22. The runtime imports shared `evaluation_pipeline/` and `postprocess/` modules from the full repository.
  23. Reuse a complete checkout or clone `https://github.com/SeerRay-Lab/Xiaomi-OCR-0.git`
  24. into a user-owned code directory. Run `skills/xiaomi-ocr/install.sh` from that checkout;
  25. do not run a detached Skill copy without setting `XIAOMI_OCR_REPO` to the checkout.
  26. For Codex use `bash install.sh --agent=codex`; for other agents use `bash install.sh`
  27. and merge its generated configuration into only the current agent.
  28. Before running `install.sh`, explain that it creates a private `.venv`, downloads the
  29. MCP dependencies listed in `requirements.txt`, and registers one MCP server entry in
  30. Codex config when applicable. Ask the user to authorize that step if they have not already
  31. authorized environment creation and dependency installation. Preserve all other MCP entries.
  32. Do not install SGLang/vLLM, create an inference environment, or download model weights
  33. unless the user's setup authorization covers those changes. State clearly if an engine or checkpoint
  34. would need to be downloaded; reuse existing compatible installations whenever possible.
  35. Keep any inference listener bound to loopback and do not claim inference is ready until an
  36. image request succeeds.
  37. For a complete agent-assisted installation, include region mode in the setup plan: explain
  38. that PaddlePaddle, PaddleX and PP-DocLayoutV3 weights are needed, then install them in the
  39. MCP environment under the user's existing setup authorization. Do not ask again when
  40. that authorization already covers these dependencies and downloads. Follow step 5 of the
  41. runtime Skill to choose a compatible CPU/GPU build, install `requirements-paddlex.txt`,
  42. and verify a real `mode="region"` request. Respect requests for page-only setup. If the
  43. platform is unsupported or installation fails, report region mode as unavailable and
  44. explain which base functions passed verification; do not claim full setup is complete.
  45. For the runtime workflow, task selection, and output handling, follow
  46. [`skills/xiaomi-ocr/SKILL.md`](skills/xiaomi-ocr/SKILL.md). The model is served locally
  47. through SGLang, vLLM, or the optional Ollama backend; the MCP service performs task-specific prompting, page/region
  48. preprocessing, and result assembly.
  49. Choose `ocr_image`/`ocr_pdf` for document parsing, `kie_image` for JSON key fields, and
  50. `vqa_image` for questions about a document image. Prefer region mode for printed documents;
  51. use whole-page mode for scene text, handwriting, calligraphy, historical documents, and
  52. other nonstandard layouts where region cuts may damage context.
README.md · README · 278 lines
  1. <div align="center">
  2. <p><img src="demo/static/img/xiaomi-logo.svg" alt="Xiaomi" width="88" height="88"></p>
  3. <h1 align="center">Xiaomi-OCR-0</h1>
  4. <p><b>A unified 0.8B model for document parsing and OCR-related understanding.</b></p>
  5. <p><b>English</b> · <a href="README_zh.md">简体中文</a></p>
  6. <p>
  7. <a href="https://huggingface.co/SeerRay-Lab/Xiaomi-OCR-0"><img src="https://img.shields.io/badge/Hugging%20Face-Model-FFD21E?logo=huggingface&amp;logoColor=FFD21E" alt="Hugging Face Model"></a>
  8. <a href="https://arxiv.org/abs/2609.36136"><img src="https://img.shields.io/badge/arXiv-2609.36136-b31b1b.svg" alt="arXiv:2609.36136"></a>
  9. <a href="https://huggingface.co/spaces/SeerRay-Lab/Xiaomi-OCR-0"><img src="https://img.shields.io/badge/Project-Page-FF6900?logo=huggingface&amp;logoColor=FFD21E" alt="Project Page"></a>
  10. <a href="https://github.com/SeerRay-Lab/Xiaomi-OCR-0"><img src="https://img.shields.io/badge/GitHub-SeerRay--Lab%2FXiaomi--OCR--0-181717?logo=github" alt="GitHub"></a>
  11. <a href="https://ollama.com/longwayxu/xiaomi-ocr-0:bf16"><img src="https://img.shields.io/badge/Ollama-BF16%20%28experimental%29-000000?logo=ollama&amp;logoColor=white" alt="Ollama BF16 (experimental)"></a>
  12. <a href="https://www.modelscope.cn/models/SeerRay-Lab/Xiaomi-OCR-0"><img src="https://img.shields.io/badge/ModelScope-Model-624AFF?logo=ModelScope&amp;logoColor=white" alt="ModelScope Model"></a>
  13. </p>
  14. </div>
  15. Xiaomi-OCR-0 is a unified 0.8B vision-language model for document parsing and OCR-centric understanding. Starting from Qwen3.5-0.8B, it is trained on an approximately 170M-sample OCR-centric corpus with Q-Mask text anchoring, continued pretraining (CPT), and mixed-task reinforcement learning (Mix-RL).
  16. ## 🏆 Benchmark highlights
  17. - 🥇 **95.24** on Real5-OmniDocBench, ranked **#1** in the comparison on this page.
  18. - 🥈 **96.83** on OmniDocBench v1.6 and **87.94** on Wild-OmniDocBench, both ranked **#2** in the comparisons on this page.
  19. - 📐 **97.84** on UniMER-Test (formula recognition) and **92.27 TEDS** on PubTabNet (table recognition).
  20. - ✍️ Historical handwriting and calligraphy: **73** on Chronicles-OCR Mature Scripts (Clerical, Cursive, Running, and Regular scripts), on par with Seed 2.0 Pro(72).
  21. - 🧠 Mean score of **83.2** across five OCR-centric VQA benchmarks, exceeding Qwen3.5-2B (**80.9**) with **0.8B** parameters.
  22. The document parsing benchmark results above (Real5-OmniDocBench, OmniDocBench v1.6, and Wild-OmniDocBench) all use the two-stage method. See [evaluation_pipeline](evaluation_pipeline/README.md) for the evaluation procedure.
  23. see [Key Performance](#key-performance) for the comparison tables and scope.
  24. ## 📰 News
  25. - **2026-11-10:** Try [Any2Markdown](https://github.com/SeerRay-Lab/Any2Markdown) to turn PDFs, Office documents, images, and web pages into clean Markdown with Xiaomi-OCR-0 as the default OCR model.
  26. - **2026-10-08:** Thanks to the community for supporting Xiaomi-OCR-0 and sharing new formats: [prithivMLmods’ GGUF](https://huggingface.co/prithivMLmods/Xiaomi-OCR-0-GGUF), [xzl01’s GGUF](https://github.com/xzl01/Xiaomi-OCR-0-GGUF), and [ByronLeeee’s BF16 NInfer conversion](https://huggingface.co/ByronLeeee/Xiaomi-OCR-0-Ninfer).
  27. - **2026-09-30:** Technical report published and model weights released.
  28. ## What it does
  29. - **Document parsing:** reads document pages and produces structured Markdown, with OTSL model outputs converted to HTML tables in Markdown and formulas in LaTeX.
  30. - **Key information extraction (KIE):** extracts requested fields as JSON.
  31. - **OCR-centric visual question answering (VQA):** answers questions about text and content in document images.
  32. - **Local agent tools:** an MCP service handles prompts, PDF/image preparation, optional region detection, parallel crop inference, and output assembly.
  33. - **Browser demo:** try page and region parsing through a local web interface.
  34. For printed, regular documents, region parsing can detect and process regions concurrently. For scene text, handwriting, calligraphy, historical books, and irregular layouts, use whole-page parsing because incorrect region boundaries can lose context.
  35. ## Repository contents
  36. | Path | Contents |
  37. |:--|:--|
  38. | [`skills/xiaomi-ocr/`](skills/xiaomi-ocr/) | Agent Skill and MCP service for OCR, PDF parsing, KIE, and VQA |
  39. | [`demo/`](demo/) | Local browser demo and sample cases |
  40. | [`example_pics/`](example_pics/) | Example inputs, reference Markdown, and demo animations mirrored from the Hugging Face model repository |
  41. | [`evaluation_pipeline/`](evaluation_pipeline/) | Reproducible e2e / two-stage inference and evaluation |
  42. | [`postprocess/`](postprocess/) | Shared table conversion, formula/text assembly and repetition handling |
  43. | [Hugging Face model](https://huggingface.co/SeerRay-Lab/Xiaomi-OCR-0) | Model card, examples, and benchmark assets |
  44. | [ModelScope model](https://www.modelscope.cn/models/SeerRay-Lab/Xiaomi-OCR-0) | Model card, examples, and benchmark assets |
  45. | [Hugging Face Space](https://huggingface.co/spaces/SeerRay-Lab/Xiaomi-OCR-0/tree/main) | Space assets and project-page source |
  46. | [`examples/`](examples/) | MCP configuration example |
  47. ## Installation
  48. Choose one of two setup paths:
  49. - [Try it with an agent](#try-it-with-an-agent): let your agent install the local model and configure the MCP service.
  50. - [Manual installation](#manual-installation): get the code, choose an inference runtime, and run the browser demo yourself.
  51. ### Try it with an agent
  52. > ⚠️ **Notice:** With your authorization, this Skill may create a Python environment, install dependencies, download model weights, install and configure the selected inference runtime (SGLang, vLLM, or Ollama), and register an MCP server. Model and layout downloads can be large. If you do not agree to this setup flow, do not execute the command below.
  53. Copy and send this to your agent:
  54. ```text
  55. Read and execute https://raw.githubusercontent.com/SeerRay-Lab/Xiaomi-OCR-0/main/SKILL.md
  56. ```
  57. The model will run on your machine. For a complete setup, the agent also installs the
  58. region-mode dependencies (PaddlePaddle, PaddleX and PP-DocLayoutV3) under your setup
  59. authorization and verifies region parsing. You can request page-only setup. If your
  60. platform does not support the layout runtime, the agent will explain the limitation.
  61. ### Manual installation
  62. Python 3.10 or newer is required. Get the repository first, then choose **one** runtime: [SGLang / vLLM](#sglang--vllm) or [Ollama](#ollama-experimental). To use region parsing, also install the [optional layout dependencies](#optional-region-mode-dependencies).
  63. For manual MCP setup, see [INSTALL.md](INSTALL.md); Ollama-specific setup is covered in [OLLAMA.md](OLLAMA.md).
  64. #### Get the repository
  65. ```bash
  66. git clone https://github.com/SeerRay-Lab/Xiaomi-OCR-0.git
  67. cd Xiaomi-OCR-0
  68. ```
  69. #### SGLang / vLLM
  70. You can run the browser demo without installing the Agent Skill or MCP server. The demo sends requests to a local SGLang or vLLM inference server at `http://127.0.0.1:8000/v1`.
  71. ##### 1. Install an inference runtime
  72. If a compatible local server is already running at the address above, skip to step 3. Otherwise, choose **one** runtime and follow its installation guide for your operating system, GPU, and driver. These runtimes have hardware-specific requirements; the linked guides list supported platforms and installation options:
  73. - [SGLang installation guide](https://docs.sglang.ai/get_started/install.html) · [supported models](https://docs.sglang.io/docs/supported-models/multimodal_language_models.md)
  74. - [vLLM GPU installation guide](https://docs.vllm.ai/en/latest/getting_started/installation/gpu/) · [Qwen3.5 model support](https://docs.vllm.ai/en/latest/api/vllm/model_executor/models/qwen3_5/)
  75. Use a release that supports the checkpoint architecture `Qwen3_5ForConditionalGeneration`. For macOS and non-NVIDIA hardware, check the runtime's platform-specific support before installing; the default commands below are for a compatible GPU installation.
  76. ##### 2. Start Xiaomi-OCR-0
  77. Run **one** of the following commands in a terminal. On first launch, the runtime downloads the checkpoint from Hugging Face if it is not already cached. Keep this terminal open while using the demo.
  78. SGLang:
  79. ```bash
  80. python -m sglang.launch_server --model-path SeerRay-Lab/Xiaomi-OCR-0 \
  81. --host 127.0.0.1 --port 8000 --context-length 16384
  82. ```
  83. vLLM:
  84. ```bash
  85. vllm serve SeerRay-Lab/Xiaomi-OCR-0 \
  86. --host 127.0.0.1 --port 8000 --max-model-len 16384
  87. ```
  88. Wait for the server to finish loading the model before continuing.
  89. ##### 3. Start the browser demo
  90. Open a second terminal in the repository directory and run:
  91. ```bash
  92. python3 -m venv .venv
  93. source .venv/bin/activate
  94. python -m pip install -r requirements.txt
  95. python demo/server.py --port 8787
  96. ```
  97. Open <http://127.0.0.1:8787>. The interface supports English and Chinese; switch in the header or open <http://127.0.0.1:8787/?lang=en> for English. The same browser page supports whole-page/region document parsing, KIE, VQA, and PDF parsing. Install the shared requirements above, including Pillow and `pypdfium2` for images and PDFs. Region mode optionally needs PaddlePaddle/PaddleX and PP-DocLayoutV3 weights. See [demo/README.md](demo/README.md) for dependencies and [INSTALL.md](INSTALL.md) for MCP setup.
  98. #### Ollama (experimental)
  99. The [community BF16 GGUF package](https://ollama.com/longwayxu/xiaomi-ocr-0:bf16) includes the text model and vision projector (about 1.8 GB). Install [Ollama](https://ollama.com/download) and keep its local service running, then pull the model:
  100. ```bash
  101. ollama pull longwayxu/xiaomi-ocr-0:bf16
  102. ```
  103. Use the repository's **MCP service or browser Demo** for OCR. They automatically apply the image preprocessing, non-thinking prompt and decoding settings used in our alignment checks; pulling the model alone does not configure these steps. No additional wrapper script is needed when using these integrations.
  104. For the browser Demo, run from the repository root with Python 3.10+:
  105. ```bash
  106. python3 -m venv .venv
  107. source .venv/bin/activate
  108. python -m pip install -r requirements.txt
  109. export XIAOMI_OCR_BACKEND=ollama
  110. export XIAOMI_OCR_MODEL=longwayxu/xiaomi-ocr-0:bf16
  111. export XIAOMI_OCR_LOCAL_URL=http://127.0.0.1:11434/v1
  112. python demo/server.py --port 8787
  113. ```
  114. Open <http://127.0.0.1:8787>. Whole-page OCR, PDF parsing, KIE and VQA share the existing pipeline. Region mode additionally requires the layout dependencies described below. See [Ollama setup and validation notes](OLLAMA.md) for direct MCP installation and conversion details.
  115. **Validation scope:** on Apple Silicon with Ollama 0.35.0, all 7 sample input token counts matched the HF BF16 reference; 5 of 7 raw responses matched exactly, with formula differences in the other two. These checks do not establish general accuracy parity or validate the benchmark scores below for Ollama.
  116. #### Optional: region-mode dependencies
  117. Region mode additionally needs PaddlePaddle, PaddleX and PP-DocLayoutV3. In the Python
  118. environment that runs the Demo or MCP server, follow the official
  119. [PaddlePaddle installation guide](https://paddlepaddle.github.io/PaddleX/latest/en/installation/paddlepaddle_install.html)
  120. to select a compatible CPU/GPU build, then the
  121. [PaddleX installation guide](https://paddlepaddle.github.io/PaddleX/latest/en/installation/installation.html).
  122. From the repository root, install the layout dependencies:
  123. ```bash
  124. python -m pip install -r requirements-paddlex.txt
  125. ```
  126. This includes `paddlex[cv]` and `shapely`; PaddlePaddle is installed separately using the
  127. official guide. PP-DocLayoutV3 weights download on first use if not cached. For CPU layout,
  128. set `XIAOMI_OCR_LAYOUT_DEVICE=cpu` in the Demo or MCP environment. Whole-page OCR, KIE and
  129. VQA work without these extra dependencies.
  130. ## Output length and task status
  131. The default output limit for document parsing and KIE is **4096 tokens**. In the
  132. Demo, open **Settings → Maximum output tokens**, enter a value, and save before
  133. running the task. With MCP, pass `max_tokens` to `ocr_image`, `ocr_pdf`, or `kie_image`.
  134. For example, these `ocr_pdf` arguments allow up to 8192 output tokens per region:
  135. ```json
  136. {"pdf_path": "/absolute/path/document.pdf", "mode": "region", "max_tokens": 8192}
  137. ```
  138. Page mode applies the limit to each page; Region mode applies it to each detected
  139. region, including text, tables, and formulas. Each PDF page/region gets its own
  140. budget. KIE applies it to one extraction; VQA uses a fixed 1024-token limit for concise
  141. answers. Generation can finish before reaching the limit.
  142. For truncated output, increase `max_tokens` if the model context has room. Image and
  143. prompt tokens also consume context: if their sum with the requested output exceeds
  144. the server's context limit, reduce the output budget or input size instead.
  145. The Demo shows the layout/OCR devices, stage, and page/region progress. MCP provides
  146. the same task information through progress notifications and `runtime_status`;
  147. use the returned `request_id` to query a specific task. See [Demo usage](demo/README.md#use)
  148. and [MCP configuration](INSTALL.md#mcp-tasks) for details.
  149. ## Community model conversions
  150. ### GGUF quantizations
  151. [prithivMLmods/Xiaomi-OCR-0-GGUF](https://huggingface.co/prithivMLmods/Xiaomi-OCR-0-GGUF) provides `BF16`, `Q3_K_L`, `Q3_K_M`, `Q4_K_M`, `Q4_K_S`, `Q5_K_M`, `Q5_K_S` and `Q6_K` weights. Image inference also requires the repository's `Xiaomi-OCR-0.mmproj-bf16.gguf` vision projector. See the [file list](https://huggingface.co/prithivMLmods/Xiaomi-OCR-0-GGUF/tree/main) for downloads and sizes. We have not evaluated these community GGUF conversions; the original model's benchmark scores and Ollama alignment checks do not establish their accuracy.
  152. [xzl01/Xiaomi-OCR-0-GGUF](https://github.com/xzl01/Xiaomi-OCR-0-GGUF) also provides Q4_K_M and Q8_0 llama.cpp weights with an F16 vision projector. See the repository for its conversion scripts and measured results.
  153. ### NInfer BF16
  154. [ByronLeeee/Xiaomi-OCR-0-Ninfer](https://huggingface.co/ByronLeeee/Xiaomi-OCR-0-Ninfer) provides a BF16 NInfer package with language and vision weights, tokenizer, and image-processing configuration. See its model page for the build guide and contributor-reported runtime and output comparisons. We have not independently reproduced those measurements; they are separate from the benchmark results for the original checkpoint above.
  155. ## Key Performance
  156. ![Parameter efficiency on OmniDocBench v1.6](https://huggingface.co/SeerRay-Lab/Xiaomi-OCR-0/resolve/main/assets/omni-parameters.png)
  157. ![Parameter efficiency on OCR-oriented VQA](https://huggingface.co/SeerRay-Lab/Xiaomi-OCR-0/resolve/main/assets/vqa-parameters.png)
  158. ![Parameter efficiency on Real5-OmniDocBench](https://huggingface.co/SeerRay-Lab/Xiaomi-OCR-0/resolve/main/assets/real5-parameters.png)
  159. ![Parameter efficiency on Wild-OmniDocBench](https://huggingface.co/SeerRay-Lab/Xiaomi-OCR-0/resolve/main/assets/wild-parameters.png)
  160. ### Document parsing: selected comparisons
  161. | Model | Size | OmniDocBench v1.6 ↑ | Real5 ↑ | Wild ↑ |
  162. |:--|--:|--:|--:|--:|
  163. | **Xiaomi-OCR-0** | **0.8B** | **96.83** | **95.24** | **87.94** |
  164. | TeleOCR | 1.2B | 96.87 | — | 88.53 |
  165. | OvisOCR2 | 0.8B | 96.58 | 92.29 | 87.91 |
  166. | PaddleOCR-VL-1.6 | 0.9B | 96.33 | 93.19 | 87.36 |
  167. | MinerU2.5-Pro | 1.2B | 95.75 | 88.94 | 87.33 |
  168. | GLM-OCR | 0.9B | 95.22 | 90.32 | 85.08 |
  169. All three columns report Overall scores.
  170. ### OCR-centric visual question answering
  171. | Model | Size | DocVQA | InfoVQA | ChartQA | OCRBench | TextVQA | Mean |
  172. |:--|--:|--:|--:|--:|--:|--:|--:|
  173. | **Xiaomi-OCR-0** | **0.8B** | **93.1** | **75.1** | **84.6** | **84.6** | **78.6** | **83.2** |
  174. | Qwen3.5-0.8B | 0.8B | 88.5 | 60.3 | 69.5 | 77.9 | 68.3 | 72.9 |
  175. | Qwen3.5-2B | 2B | 92.4 | 72.4 | 77.0 | 85.9 | 76.9 | 80.9 |
  176. | Qwen3.5-4B | 4B | 94.4 | 80.4 | 82.4 | 86.6 | 80.8 | 84.9 |
  177. | MiniCPM-V-4.5 | 8B | 84.9 | 69.6 | 87.4 | 89.0 | 82.2 | 82.6 |
  178. Mean is the arithmetic average of the five benchmarks on a 0–100 scale.
  179. ## Reproduce the evaluation
  180. See the [evaluation guide](evaluation_pipeline/README.md) for e2e inference, vLLM
  181. parameters, PaddleX PP-DocLayoutV3 layout at threshold 0.25 for two-stage inference,
  182. and scoring with the OmniDocBench evaluation Docker image. We publish the
  183. [exact evaluation configuration](evaluation_pipeline/configs/end2end.yaml) used in our runs.
  184. ## Citation
  185. If you follow our work, please cite the following:
  186. ```bibtex
  187. @misc{chen2026xiaomiocr0technicalreport,
  188. title={Xiaomi-OCR-0 Technical Report},
  189. author={Xin Chen and Anan Du and Feng Feng and Pei Fu and Jian Luan and Longwei Xu and Shaojie Zhang and Hang Li and Heng Qu and Cheng Tan},
  190. year={2026},
  191. eprint={2609.36136},
  192. archivePrefix={arXiv},
  193. primaryClass={cs.CV},
  194. url={https://arxiv.org/abs/2609.36136},
  195. }
  196. ```

Ratings and comments

Sign in to rate this skill.

Sign in

Free to use. If it’s helped, pass it on.