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).

• 🥇 95.24 on Real5-OmniDocBench, ranked #1 in the comparison on this page.
• 🥈 96.83 on OmniDocBench v1.6 and 87.94 on Wild-OmniDocBench, both ranked #2 in the comparisons on this page.
• 📐 97.84 on UniMER-Test (formula recognition) and 92.27 TEDS on PubTabNet (table recognition).
• ✍️ Historical handwriting and calligraphy: 73 on Chronicles-OCR Mature Scripts (Clerical, Cursive, Running, and Regular scripts), on par with Seed 2.0 Pro(72).
• 🧠 Mean score of 83.2 across five OCR-centric VQA benchmarks, exceeding Qwen3.5-2B (80.9) with 0.8B parameters.

The document parsing benchmark results above (Real5-OmniDocBench, OmniDocBench v1.6, and Wild-OmniDocBench) all use the two-stage method. See evaluation_pipeline for the evaluation procedure.

see Key Performance for the comparison tables and scope.

• 2026-11-10: Try Any2Markdown to turn PDFs, Office documents, images, and web pages into clean Markdown with Xiaomi-OCR-0 as the default OCR model.
• 2026-10-08: Thanks to the community for supporting Xiaomi-OCR-0 and sharing new formats: prithivMLmods’ GGUF, xzl01’s GGUF, and ByronLeeee’s BF16 NInfer conversion.
• 2026-09-30: Technical report published and model weights released.

• Document parsing: reads document pages and produces structured Markdown, with OTSL model outputs converted to HTML tables in Markdown and formulas in LaTeX.
• Key information extraction (KIE): extracts requested fields as JSON.
• OCR-centric visual question answering (VQA): answers questions about text and content in document images.
• Local agent tools: an MCP service handles prompts, PDF/image preparation, optional region detection, parallel crop inference, and output assembly.
• Browser demo: try page and region parsing through a local web interface.

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.

Choose one of two setup paths:

• Try it with an agent: let your agent install the local model and configure the MCP service.
• Manual installation: get the code, choose an inference runtime, and run the browser demo yourself.

⚠️ 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.

Python 3.10 or newer is required. Get the repository first, then choose one runtime: SGLang / vLLM or Ollama. To use region parsing, also install the optional layout dependencies.

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

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

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.

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 · supported models
• vLLM GPU installation guide · Qwen3.5 model support

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.

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.

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

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 for dependencies and INSTALL.md for MCP setup.

The community BF16 GGUF package includes the text model and vision projector (about 1.8 GB). Install Ollama and keep its local service running, then pull the model:

ollama pull longwayxu/xiaomi-ocr-0:bf16

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.

For the browser Demo, run from the repository root with Python 3.10+:

python3 -m venv .venv
source .venv/bin/activate
python -m pip install -r requirements.txt
export XIAOMI_OCR_BACKEND=ollama
export XIAOMI_OCR_MODEL=longwayxu/xiaomi-ocr-0:bf16
export XIAOMI_OCR_LOCAL_URL=http://127.0.0.1:11434/v1
python demo/server.py --port 8787

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 for direct MCP installation and conversion details.

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.

Region mode additionally needs PaddlePaddle, PaddleX and PP-DocLayoutV3. In the Python environment that runs the Demo or MCP server, follow the official PaddlePaddle installation guide to select a compatible CPU/GPU build, then the PaddleX installation guide. From the repository root, install the layout dependencies:

python -m pip install -r requirements-paddlex.txt

This includes paddlex[cv] and shapely; PaddlePaddle is installed separately using the official guide. PP-DocLayoutV3 weights download on first use if not cached. For CPU layout, set XIAOMI_OCR_LAYOUT_DEVICE=cpu in the Demo or MCP environment. Whole-page OCR, KIE and VQA work without these extra dependencies.

The default output limit for document parsing and KIE is 4096 tokens. In the Demo, open Settings → Maximum output tokens, enter a value, and save before running the task. With MCP, pass max_tokens to ocr_image, ocr_pdf, or kie_image. For example, these ocr_pdf arguments allow up to 8192 output tokens per region:

{"pdf_path": "/absolute/path/document.pdf", "mode": "region", "max_tokens": 8192}

Page mode applies the limit to each page; Region mode applies it to each detected region, including text, tables, and formulas. Each PDF page/region gets its own budget. KIE applies it to one extraction; VQA uses a fixed 1024-token limit for concise answers. Generation can finish before reaching the limit.

For truncated output, increase max_tokens if the model context has room. Image and prompt tokens also consume context: if their sum with the requested output exceeds the server's context limit, reduce the output budget or input size instead.

The Demo shows the layout/OCR devices, stage, and page/region progress. MCP provides the same task information through progress notifications and runtime_status; use the returned request_id to query a specific task. See Demo usage and MCP configuration for details.

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 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.

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.

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.

All three columns report Overall scores.

Mean is the arithmetic average of the five benchmarks on a 0–100 scale.

See the evaluation guide for e2e inference, vLLM parameters, PaddleX PP-DocLayoutV3 layout at threshold 0.25 for two-stage inference, and scoring with the OmniDocBench evaluation Docker image. We publish the exact evaluation configuration used in our runs.

If you follow our work, please cite the following:

@misc{chen2026xiaomiocr0technicalreport,
      title={Xiaomi-OCR-0 Technical Report},
      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},
      year={2026},
      eprint={2609.36136},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2609.36136},
}