Running Gemma 4 locally with LM Studio's new headless CLI and Claude Code | Hacker News

Mixture of Experts models still require all weights to be loaded into memory, even if fewer experts are used per forward pass.

Consulting fewer experts can raise token throughput but does not by itself reduce peak VRAM usage.

Using an inference engine that offloads some experts to CPU RAM can lower VRAM requirements and let larger MoE models run on smaller GPUs.

Claude Code is currently a popular frontend for local model hosting among programmers, often used with tools like LM Studio.