K-Quantization

A mixed-precision matrix quantization method implemented in llama.cpp that compresses model weights by using different bit widths for different layers, significantly reducing VRAM usage while preserving inference capability.