README.md

metadata

language:
  - en
  - zh
tags:
  - vision
  - ocr
  - deepseek
  - quantization
  - int4
license: other
base_model: deepseek-ai/DeepSeek-OCR

DeepSeek-OCR INT4 Merged Model

This is a merged version of the DeepSeek-OCR model that was quantized to INT4 and then dequantized back to full precision.

Model Details

• Base Model: deepseek-ai/DeepSeek-OCR
• Quantization: INT4 (during storage)
• Final Format: Full precision (FP32/FP16)
• Source: Quantized weights from SamMikaelson/deepseekocr-randomreal

Usage

from transformers import AutoModelForCausalLM, AutoProcessor
from PIL import Image

# Load model and processor
model = AutoModelForCausalLM.from_pretrained(
    "YOUR_USERNAME/OCR-int4-merged",
    trust_remote_code=True
)
processor = AutoProcessor.from_pretrained(
    "YOUR_USERNAME/OCR-int4-merged",
    trust_remote_code=True
)

# Process image
image = Image.open("document.jpg")
inputs = processor(images=image, return_tensors="pt")

# Generate OCR output
outputs = model.generate(**inputs, max_new_tokens=512)
text = processor.decode(outputs[0], skip_special_tokens=True)
print(text)

Notes

This model was created by:

1. Quantizing the original DeepSeek-OCR to INT4
2. Dequantizing back to full precision
3. Merging into a standard model format

The process allows for efficient storage and distribution while maintaining full model capabilities.