---
title: GLM-OCR Pruned 8-bit Safetensors (1.3GB)
emoji: 🚀
license: mit
language:
- en
- fr
- es
- ru
- de
- ja
- ko
- zh
base_model:
- zai-org/GLM-OCR
pipeline_tag: image-text-to-text
library_name: transformers
tags:
- pruning
- bitsandbytes
- int8
---

# GLM-OCR-Pruned-8bit
![Model Size](https://img.shields.io/badge/Disk-1.3GB-brightgreen) ![GPU](https://img.shields.io/badge/GPU-2.3GB-blue) ![Quant](https://img.shields.io/badge/8--bit-✅-orange)

**Production GLM-OCR: 52% smaller (2.7GB→1.3GB), fully 8-bit, OCR optimized**

## 📊 Performance
| Metric | Original | **Optimized** |
|--------|----------|---------------|
| **Parameters** | 1.1B | **1.1B (4.3% pruned)** |
| **Disk** | 2.7GB | **1.3GB** (52%↓) |
| **GPU** | 3.5GB+ | **2.3GB** |
| **Speed** | 1x | **2-3x** |

## 🚀 Quickstart
```python
from transformers import BitsAndBytesConfig, AutoProcessor, AutoModelForImageTextToText
import torch

MODEL_PATH = "ManiKumarAdapala/glm-ocr-pruned-8bit"

messages = [
    {
        "role": "user",
        "content": [
            {
                "type": "image",
                "url": "Image.jpeg"
            },
            {
                "type": "text",
                "text": "Text Recognition:"
            }
        ],
    }
]

quant_config = BitsAndBytesConfig(load_in_8bit=True)

processor = AutoProcessor.from_pretrained(MODEL_PATH)
model = AutoModelForImageTextToText.from_pretrained(
    pretrained_model_name_or_path=MODEL_PATH,
    quantization_config=quant_config,
    device_map="auto",
)

inputs = processor.apply_chat_template(
    messages,
    tokenize=True,
    add_generation_prompt=True,
    return_dict=True,
    return_tensors="pt"
).to(model.device)

inputs.pop("token_type_ids", None)

generated_ids = model.generate(**inputs, max_new_tokens=8192)

output_text = processor.decode(generated_ids[0][inputs["input_ids"].shape[1]:], skip_special_tokens=False)

print(output_text)
```

## 🛠 Optimizations Applied

- ✅ Selective Pruning: q_proj, v_proj, fc2, vision_tower (52%)
- ✅ BitsAndBytes 8-bit: Linear8bitLt (vision+text decoder)
- ✅ Protected: lm_head, early vision, final decoder layers


## 📚 Citation

```bibtex
@misc{GLM-OCR-Pruned8bit-2026,
  author = {Mani, {ADAPALA MANI KUMAR} and {ZAI-org}},
  title = {GLM-OCR Pruned & 8-bit quantized (1.1B params, 4.3% sparsity)},
  year = {2026},
  month = {march},
  publisher = {Hugging Face},
  url = {https://huggingface.co/adapala-manikumar/glm-ocr-pruned-8bit},
  note = {1.3GB disk, 2.3GB GPU, OCR optimized, MIT}
}
```

<font size="2">
  
 **Acknowledgements (from ZAI-org/GLM-OCR)**

This project is inspired by the excellent work of:
- [PP-DocLayout-V3](https://huggingface.co/PaddlePaddle/PP-DocLayoutV3) (Apache 2.0)
- [PaddleOCR](https://github.com/PaddlePaddle/PaddleOCR) 
- [MinerU](https://github.com/opendatalab/MinerU)

**License Notice**: The GLM-OCR model is MIT licensed. When using the complete OCR pipeline, users should comply with Apache License 2.0 for PP-DocLayoutV3 components.
</font>