A llama.cpp PR makes Q2_0 3.0–3.6x faster on x86 CPUs, 8B decode goes 2.39 → 8.20 tok/s

Collecting trend data

The percentage is based on available heat signal, not comment count or independent people.