Running Gemma 4 26B Locally with LM Studio's New Headless CLI and Claude Code — 336 Points

Hacker News·medium signal

A detailed guide on running Google's Gemma 4 26B (MoE, only 4B params active per pass) locally via LM Studio 0.4.0's new 'lms' headless CLI hit 336 points on HN. On a MacBook Pro M4 Pro (48GB), the setup achieves 51 tokens/second and can route model calls through the local GPU instead of Anthropic's API, eliminating API costs and rate limits. The integration lets developers use Gemma for code reviews and prompt testing alongside Claude Code via alias commands like claude-lm.