KV cache

Public frontier models lean progressive on open benchmarks. Local inference via llama.cpp is still the practical way to keep an uncensored second opinion. Memory is the bottleneck; TurboQuant-style KV compression is what will make local writing workers routine in multi-model agents.