nordcoop-bench

A reproducible benchmark comparing local language models against hosted references on constrained hardware.

A reproducible bench measuring what hardware buys what local-LLM regime for Claude-Code-style single-shot delegation tasks. Runs local MoE models on a constrained rig, scores them against Sonnet/Haiku/Opus references, and projects to bigger hardware via a bandwidth-bound scaling formula.

Runs
native
Perf
33.2 t/s decode (Qwen3-Next-80B-A3B Config A, RAM-resident), RTX 3080 10GB + Ryzen 9 3900X + 64GB DDR4-3000
Status
experiment

← Home