With the price of tokens set to soar, and coding with AI now the norm, running local models is becoming increasingly important to minimise costs.

I use Claude Code on a Windows machine as my primary development tool, with Gemini Flash for reviews. However, if I do start hitting limits, I use local models for simpler tasks.

I have a Mac Mini M4 with 24GB of RAM, but hosting and running coding models is not completely straight-forward. The following combinations have worked for me:

oMLX, Claude Code, Gemma4

Software

• Server: oMLX running on my Mac Mini
• Model: gemma-4-E4B-it-MLX-4bit
• Coding Tool: Claude Code (configured per below)

Startup Script

If I want to switch to a local model, I use this script to configure Claude in a single console session.

$model = "gemma-4-E4B-it-MLX-4bit"

# Redirect Claude Code to your local Ollama server
$env:ANTHROPIC_BASE_URL = "http://172.168.1.74:8000"
$env:ANTHROPIC_API_KEY = ""
$env:ANTHROPIC_AUTH_TOKEN = ""

# Map Claude Code's model tier requests to your local model name
$env:ANTHROPIC_DEFAULT_SONNET_MODEL = $model
$env:ANTHROPIC_DEFAULT_HAIKU_MODEL = $model
$env:ANTHROPIC_DEFAULT_OPUS_MODEL = $model
$env:ANTHROPIC_MODEL = $model
$env:API_TIMEOUT_MS = "3000000"
$env:CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC = "1"

# launch Claude
claude