Xiaomi's 0.8B OCR-specific vision-language model (trained from Qwen3.5-0.8B-Base), a single model that simultaneously performs full-page document parsing (outputting Markdown, with tables in OTSL, formulas in LaTeX, and content in reading order), single-region text/table/formula recognition, OCR-based question answering, and key information extraction according to JSON Schema.
Upvotes 62 · Likes 115 · Downloads 15472
Reasons
- Officially claimed to rank first on Real5-OmniDocBench 95.24, and second on both OmniDocBench v1.6 96.83 and Wild-OmniDocBench 87.94 (behind only the 1.2B TeleOCR 96.87 / 88.53); formula recognition on UniMER-Test 97.84, table recognition on PubTabNet 92.27.
- Average score of 83.2 across five OCR question-answering benchmarks, higher than Qwen3.5-2B's 80.9 but lower than Qwen3.5-4B's 84.9; score of 73 on Chronicles-OCR for historical handwriting and calligraphy (clerical, cursive, running, and regular scripts), comparable to Seed 2.0 Pro's 72; specifically trained for documents that are photographed, distorted, or tilted.
- Built with approximately 1.7 billion OCR data points, constructed via automated annotation processes including cross-voting among multiple OCR models and verification by rendering back to images; post-training uses mixed-task reinforcement learning; released with the technical report on 9 30, not included in previous editions (MiniCPM-OCR-Jev, included yesterday, lists it as one of the scorable parsers).
Usage
- Transformers: uv pip install "transformers==5.17.0" "torch==2.14.0" "torchvision==0.29.0" pillow, load with AutoModelForImageTextToText, set do_sample=False and max_new_tokens=4096; runs on CUDA, Apple mps, and CPU.
- Change prompts by task: use the long prompt "Extract all information from the main body…" from the card for full pages; use "Extract the text in the image." / "Parse the table in the image into OTSL." / convert formulas to LaTeX for single regions; use "Extract key information in the image" followed by your own JSON Schema for field extraction.
- Ollama (experimental): ollama pull longwayxu/xiaomi-ocr-0:bf16 (approx. 1.8 GB), then connect to the repository's built-in MCP service or browser demo according to OLLAMA.md on GitHub; the two-stage process (PP-DocLayoutV3) involving layout analysis followed by region-by-region recognition, and batch processing scripts, are available in the GitHub repository.
Risks
- All scores are official summaries, and comparison values come from respective reports; community-converted versions such as GGUF and NInfer have not been officially evaluated.
- Table output is in OTSL format; converting to HTML / Markdown tables requires manual handling; pulling the model via Ollama alone does not automatically include the official preprocessing and decoding settings.
- The code in the card still references the old repository name SeerRay-Lab/Xiaomi-OCR-0 (GitHub will redirect to the current repository); VRAM requirements are not specified.