---
license: apache-2.0
language:
- ar
base_model: Qwen/Qwen3-VL-8B
pipeline_tag: image-to-text
tags:
- gguf
- llama.cpp
- qwen3_vl
- vision-language
- ocr
- document-understanding
- information-extraction
- arabic
- legal-documents
- egypt
---

# Qwen3-VL-8B Contracts OCR (GGUF)

This repository provides **GGUF quantizations** of **Qwen3-VL-8B-Contracts-OCR**, a fine-tuned vision-language model specialized in extracting structured information from scanned Egyptian company incorporation contracts written in Arabic.

> **The full Transformers version, model card, usage example, training details, and documentation here:**
>
> 👉 **[Qwen3-VL-8B-Contracts-OCR](https://huggingface.com/a-mo-yehia/Qwen3-VL-8B-Contracts-OCR)**


---

# About

These GGUF files are intended for inference with:

- llama.cpp
- LM Studio (when vision support is available)
- llama-server
- compatible GGUF runtimes

The original fine-tuned model was trained using **Unsloth** on top of **Qwen3-VL-8B**.

---

# Available Quantizations

| File | Description |
|------|-------------|
| `model_f16.gguf` | Full FP16 precision |
| `model_q8_0.gguf` | 8-bit quantization |
| `model_q4_k_m.gguf` | Recommended balance between quality and speed |
| `mmproj_f16.gguf` | Vision projector (required for image inference) |

---

# Which Quantization Should I Use?

| Quantization | Recommendation |
|--------------|---------------|
| Q4_K_M | Best overall choice for most users |
| Q8_0 | Better quality with higher memory usage |
| F16 | Maximum quality and accuracy |

---

# Requirements

For vision inference you **must** load:

- one GGUF model
- `mmproj_f16.gguf`

The vision projector is required for processing images.

---

# Model Purpose

The model extracts structured information from scanned Egyptian company incorporation contracts, including:

- Company name
- Company type
- Business activities
- Company address
- Capital information
- Partners and shareholders
- Contract articles
- Tables (returned as Markdown)
- Other legal fields

The output is a structured JSON document.

---

# Language

Arabic

---

# License

Apache-2.0