MoE expert expansion ported to llama.cpp: route more experts than native top-K at runtime, no training

Specific-Tax-6700 · reddit · 2026-09-07

A developer proposed a training-free inference architecture for sparse MoE models and ported expert expansion to llama.cpp: running MoE models with more routed experts than the native top-K (8→x), with an adaptive threshold, 99→50% linear influence decay, and per-layer range control—aiming to increase active parameters for more succinct reasoning without any training or fine-tuning.

The change is runtime-only and works across all backends. Tested on Qwen 3.6 35B A4B+; docs and code are on the moe-expansion branch of a llama.cpp fork on GitHub.

Related event: Dev builds llama.cpp fork enabling MoE expert expansion(2 posts)→

Original post →

More from Infra

Infra channel →