OvisOCR2-GGUF

OvisOCR2 - GGUF Quantizations

<p align="center"> <img src="https://cdn-uploads.huggingface.co/production/uploads/658a8a837959448ef5500ce5/vRCIu5QD8VuIJolkC_ZHQ.png" alt="Ovis" width="30%" /> </p>

This repository contains GGUF format quantizations of OvisOCR2, a compact 0.8B end-to-end model for page-level document parsing. The original model was developed by ATH-MaaS by post-training Qwen3.5-0.8B to parse full document pages directly into clean Markdown (including LaTeX formulas, HTML tables, and layout components).

OvisOCR2 establishes a new state-of-the-art for compact document understanding, scoring 96.58 on OmniDocBench v1.6 and outperforming traditional, multi-stage layout analysis pipelines.

---

Available Files

Main Text Models

Multimodal Projectors (mmproj)

Note: Because OvisOCR2 is a vision-language model, you must download one of these image processing units alongside your choice of the text models listed above.

• mmproj-F32.gguf (402 MB) - Unquantized full precision projector.
• mmproj-F16.gguf (205 MB) - Recommended standard performance/size option.
• mmproj-BF16.gguf (207 MB) - Target alternative precision layout.

---

Inference Guide (llama.cpp)

To run multimodal OCR tasks using these GGUF files, you need to use the llama-minicpmv-cli or llama-llava-cli tool (depending on your build version of llama.cpp) to handle simultaneous image and text tokens.

Basic Command Line Example

# Run parsing via llama.cpp cli tools
./llama-minicpmv-cli \
  -m OvisOCR2-Q5_K_M.gguf \
  --mmproj mmproj-F16.gguf \
  --image /path/to/your/document_page.jpg \
  -p "<|im_start|>user\nExtract all readable content from the image in natural human reading order and output the result as a single Markdown document. Format formulas as LaTeX. Format tables as HTML: <table>...</table>. Preserve the original text without translation.<|im_end|>\n<|im_start|>assistant\n" \
  -n 4096 \
  --temp 0.0