llama.cpp speeds up 1-bit CPU inference by 55x

This is the final piece of the puzzle for 1-bit LLMs — making them actually fast on the hardware they were meant to save.

• –55x speedup on modern CPUs transforms 1-bit models from academic curiosities into production-ready local AI tools.
• –SSSE3 support is a major win for legacy hardware, breathing new life into laptops and servers over a decade old.
• –The shift from generic scalar fallbacks to optimized SIMD kernels bridges the "performance gap" where 1-bit was paradoxically slower than 4-bit due to lack of software maturity.
• –While Apple Silicon and NVIDIA still lead, the EPYC/Xeon gains make high-density 1-bit inference commercially viable for CPU-only cloud instances.
• –This effectively standardizes the "Bonsai" 1.7B-8B architecture as the go-to for edge and low-RAM deployments.