README.md

metadata

library_name: openvino
tags:
  - openvino
  - int4
  - quantized
  - tiny-model
  - minicpm
  - visual-language-model
license: apache-2.0

Tiny MiniCPM-o-2_6 Model (6MB INT4 Quantized)

This is a tiny random version of the MiniCPM-o-2_6 model, optimized for testing purposes.

Model Details

• Model Size: ~6-7MB (INT4 quantized)
• Original Model: MiniCPM-o-2_6
• Quantization: INT4 pipeline quantization
• Vocabulary Size: 50,000 tokens (reduced from 151,700)
• Format: OpenVINO IR

Model Architecture

• Maintains MiniCPMO class compatibility
• Full INT4 quantization for all components:
  • Language model
  • Vision embeddings
  • Resampler

Usage

With Optimum-Intel

from optimum.intel import OVModelForVisualCausalLM
from transformers import AutoProcessor

model = OVModelForVisualCausalLM.from_pretrained(
    "M-Ziyo/tiny-random-MiniCPM-o-2_6-6mb",
    export=False,  # Already quantized
    trust_remote_code=True
)

processor = AutoProcessor.from_pretrained(
    "optimum-intel-internal-testing/tiny-random-MiniCPM-o-2_6",
    trust_remote_code=True
)

Validation

python validate_tiny_minicpm.py --model-path M-Ziyo/tiny-random-MiniCPM-o-2_6-6mb

Files

• OpenVINO IR files (.xml, .bin) for all model components
• Configuration files (config.json, openvino_config.json)
• Python model files for custom architecture
• Processor files

Notes

• This is a test model with random weights
• Processor should be loaded from the original model: optimum-intel-internal-testing/tiny-random-MiniCPM-o-2_6
• Compatible with Optimum-Intel test suite