LGxNDs/Geeked-Out-Quantization-Software

IQ2_M - GeekedOut Quantizer

GeekedOut Quantizer is a specialized 2-bit quantization tool that implements the IQ2_M (Intelligent Quants) scheme for efficient model compression. This repository showcases IQ2_M quantized models with extreme low-bit precision while preserving critical model capabilities through intelligent weight allocation.

About GeekedOut Quantizer

GeekedOut Quantizer is an advanced quantization framework designed to:

• Achieve 2-bit compression using the IQ2_M scheme
• Maintain high-quality inference performance
• Support GGUF format for local deployment
• Optimize memory efficiency through mixed-precision techniques

The IQ2_M Intelligence Concept

GeekedOut Quantizer models are designed with intelligence as their primary capability. Through intelligent weight allocation, intelligence is preserved in critical parameters while less important weights are packed into minimal bit formats:

• Mixed precision - different weights receive varying bit allocations based on their sensitivity and importance
• Block-wise quantization with optimized scaling factors applied across weight blocks
• 2-bit compression achieving extreme low-bit precision while preserving critical model capabilities
• Smart allocation where critical parameters are preserved in higher precision while less important weights are packed into minimal bit formats

The Quantization Process

GeekedOut uses the A:\Geeked.Out software to create models that are intelligent through:

1. Intelligent calibration - imatrix-based calibration for optimal quantization quality
3. Mixed-precision allocation - critical parameters receive higher precision while less important weights receive minimal bit formats
5. Block-wise optimization - optimized scaling factors applied across weight blocks
7. Smart allocation - intelligence is preserved through intelligent weight distribution

IQ2_M Quantization Features

The IQ2_M (Intelligent Quants) quantization scheme features:

• The quantized models retain conversational capability while achieving significant size reduction
• Compatible with llama.cpp, LM Studio, Jan, and other local inference frameworks
• Uses imatrix-based calibration for optimal quantization quality
• Developed by GeekedOut - focused on intelligent quantization methods

Supported Use Cases

GeekedOut Quantizer models are designed for:

• Conversational AI applications where intelligence is preserved through IQ2_M quantization
• Local inference with llama.cpp, LM Studio, Jan, and similar tools
• Memory-efficient deployment scenarios
• Practical everyday use cases requiring reduced memory footprint

Usage Instructions

To load IQ2_M quantized models locally using llama.cpp or compatible inference frameworks. The GGUF files are split into two parts for efficient storage (00001-of-00002 and 00002-of-00002).

Example:

# Load the IQ2_M quantized model using llama.cpp

llama.cpp -hf LGxNDs/IQ2_M-2Bit-Quantization-By-Geeked-Out-Ai

Technical Notes

• IQ2_M quantization maintains conversational capability while achieving significant size reduction
• Compatible with llama.cpp, LM Studio, Jan, and other local inference frameworks
• Uses imatrix-based calibration for optimal quantization quality
• Developed by GeekedOut - focused on intelligent quantization methods using A:\Geeked.Out software