llama.cpp hits 1-bit quantization for native Bonsai models

Native 1-bit quantization is the definitive answer to memory-constrained local LLM execution, proving that 8B models can run comfortably on standard laptop CPUs without a discrete GPU. The Q1_0 scheme uses 128-weight groups to achieve 1.125 bits per weight with near-zero KL divergence loss. Native training by PrismML avoids the typical accuracy cliff seen when quantizing pre-trained FP16 models down to 1-bit. ARM NEON optimizations provide a 4-5x speedup, making 1.07 GiB 8B models fast enough for real-time interaction on entry-level hardware. This shift favors parameter count over precision, suggesting that wider, ultra-compressed models are the future of edge AI. Integration into llama.cpp ensures immediate cross-platform support across Mac, Windows, and Linux.