---
library_name: transformers
license: apache-2.0
tags:
- vision
- multimodal
- tiny-model
- minicpm
pipeline_tag: image-to-text
---

# Tiny MiniCPM-o-2_6 Model

A minimal, optimized version of MiniCPM-o-2_6 for testing and development purposes.

## Model Details

- **Model Size**: ~54 MB (PyTorch safetensors format)
- **Format**: PyTorch safetensors (not OpenVINO IR)
- **Vocabulary Size**: 50,000 tokens (reduced from 151,700)
- **Architecture**: MiniCPM-o-2_6 with optimized dimensions

## Model Configuration

- **hidden_size**: 128 (reduced from 168)
- **intermediate_size**: 8 (reduced from 16)
- **num_hidden_layers**: 2
- **num_attention_heads**: 2 (reduced from 28)
- **query_num**: 64

## Usage

```python
from transformers import AutoProcessor, AutoModelForCausalLM
from PIL import Image

# Load processor and model
processor = AutoProcessor.from_pretrained("M-Ziyo/tiny-random-MiniCPM-o-2_6-mini", trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained("M-Ziyo/tiny-random-MiniCPM-o-2_6-mini", trust_remote_code=True)

# Prepare inputs
prompt = "<|im_start|>user\n(<image>./</image>)\nWhat is in the image?<|im_end|>\n<|im_start|>assistant\n"
image = Image.open("your_image.jpg")

inputs = processor([prompt], [image], return_tensors="pt")

# Generate
result = model.generate(**inputs, max_new_tokens=50)
decoded = processor.tokenizer.batch_decode(result[:, inputs["input_ids"].shape[1]:])
print(decoded)
```

## Model Features

- ✅ **PyTorch format** with safetensors (not OpenVINO IR)
- ✅ **Optimized size** (~54 MB vs original)
- ✅ **Weight copying** from original model for better output quality
- ✅ **Diverse output** (not just repetitive characters)

## Notes

- This is a minimal test model for development purposes
- Model weights are copied from the original model for better initialization
- Designed for testing Optimum-Intel integration

## Citation

Based on MiniCPM-o-2_6 from OpenBMB.