Proposes CPU inference method with ternary weights for 10B model at 100 tok/s

The idea suggests that CPU decode speed depends on active parameters per token, not total parameters. Aims to achieve 100 tokens/s on a mid-level PC using ternary weights and small active batch.