GitHub Repository

Run GLM-4.5-Air (110B) on a 16GB-RAM consumer machine - identify the best memory allocation to overcome standard hardware limitations in Local LLM applications. Placement beats budget. Falsification-Tested laws, probes and recipes for LLMs on commodity hardware

10 starsPython

Run GLM-4.5-Air(110B)on a 16GBRAM consumer machine

Name: Run GLM-4.5-Air(110B)on a 16GBRAM consumer machine
Availability: InStock
Author: federicoTXTS

by federicoTXTS·Jul 23, 2026·1 point·0 comments

Visit Project View on HN

AI Analysis

●●●●GemWizardryBig Brain

Runs 110B models on 16GB RAM by proving placement beats budget with measured laws.

Strengths

•Four falsification-tested laws predict decode speed from 7B to 744B parameters accurately.
•Data-free 2-bit quantization achieves better quality gaps than calibrated SOTA methods.
•Interactive calculator plots user configs against every validated measurement automatically.

Weaknesses

•Validation limited to a single 2016 GTX 1060 desktop configuration for all laws.
•Sub-1 tok/s speeds on SATA drives make interactive usage impractical for most.