# Extreme quantization optimization

MiniCPM4-8B is an efficient large language model designed specifically for edge devices. Through innovations in four dimensions: model architecture, training data, training algorithms, and inference systems, it achieves extreme efficiency improvements.

A horror-optimized version based on Google's Gemma-3 model, featuring extreme quantization technology and horror enhancement matrix, supporting a 32k context window

A horror-style instruction-tuned version based on Google's Gemma-3 model, featuring Neo Imatrix technology and extreme quantization, supporting 128k context length

An extreme quantization version based on Google's Gemma-3 model, enhanced with NEO Imatrix technology, supporting 128k context length and suitable for full-scenario tasks