Quantization, LoRA, and the 8% Problem Benchmarking Local LLMs for Production AI | Dark Hacker News