Running Google Gemma 4 Locally With LM Studio’s New Headless CLI & Claude Code
LM Studio 0.4.0 introduced llmster and the lms CLI. Here is how I set up Gemma 4 26B for local inference on macOS that can be used with Claude Code.
Technical Setup Guide
Summary
A step-by-step guide for installing Ollama on an Apple Silicon Mac mini and deploying the Gemma 4 26B model with automatic startup, preloading, and keep‑alive configuration. It covers prerequisites, installation commands, model download, testing, and creating launch agents to maintain the model in memory.
Highlights
- Installs Ollama via Homebrew cask and verifies the service is running
- Pulls the 26‑billion‑parameter Gemma 4 model (~17 GB) and confirms GPU acceleration
- Creates a macOS launch agent to preload the model on system startup
- Configures environment variables to prevent automatic model unload
- Provides commands for testing the model and monitoring resource usage
auto-generated
Context
Audience
Developers and AI engineers who want to run Gemma 4 26B locally on Apple Silicon devices
DomainMachine Learning Operations
Formatmarkdown guide with code snippets
Accessfree online via Gist
Source Overview
Gemma 4 26B is a 26-billion-parameter mixture-of-experts model from Google DeepMind that activates only 3.8 billion parameters during inference for low-latency performance.
This guide provides step-by-step instructions for running Gemma 4 26B locally on an Apple Silicon Mac mini via Ollama, including auto-start configuration, model preloading, and persistent memory management. The setup requires at least 24GB unified memory, leverages Apple's MLX framework for GPU acceleration, and exposes a local API for integration with coding agents and agentic workflows.
Topics
Related
Ollama documentationApple MLX framework overviewMac mini hardware specificationsPrompt engineering for large language models
