friend.cpp: experimental local LLM engine adds blue noise sampling and adaptive speculative decoding

amplifiedamp · x · 2026-09-23

A developer released friend.cpp, an experimental local LLM inference engine forked from koboldcpp, featuring adaptive speculative decoding, a tiered prompt cache, per-request adapter hot-swap, and steering. The author also proposes blue noise sampling: generating stable, coherent continuations from models prone to mode collapse (including base models) without reducing temperature. A fork of llama.cpp is available to try it out.

Related event: Blue noise sampling reduces mode collapse in LLMs(2 posts)→

Original post →

More from Infra

Infra channel →