arxivcs.LG2026-07-18
Half the Experts, All the Code: One-Shot Domain Pruning of Mixture-of-Experts LLMs for Coding
The strongest open-weight coding models are mixture-of-experts (MoE) networks: most of their size comes from large pools of "expert" subnetworks, of which only a few act on any token. That pool is why these models do not fit on the machines most developers own, yet for a user who…