Quantization's impact on model capabilities broken down by benchmark

Systematic tests on several models compare FP16 vs GGUF quant levels across math (GSM8K), code (HumanEval), reasoning (ARC-Challenge), and knowledge recall (MMLU-Pro). Surprising results show different capabilities degrade unevenly at various quantization levels.