Mac Mini LLM Experiment — AI skill for Claude Code

Running Local LLM Coding Agents on a 24GB Mac Mini

**Can a Mac Mini with 24GB of unified memory run an AI coding agent locally?** I spent an evening finding out. Here's everything that happened — the configs, the crashes, the infinite loops, and the lessons learned.

The Setup

• Machine: Mac Mini, Apple Silicon, 24GB unified memory
• LLM Runtime: Ollama
• Coding Agent: OpenCode — an open-source terminal-based AI coding assistant (similar to Claude Code) that supports multiple LLM backends
• Goal: Get a local model to read GitHub issues, navigate a Laravel codebase, and create pull requests — fully agentic, multi-step coding workflows

The Models I Tested

Model	Type	Model Size	Context Window	Outcome
gemma4:26b	Dense	17 GB	4k (default)	Tool names wrong (shell instead of bash)
qwen2.5-coder:14b	Dense	9 GB	4k (default)	Raw JSON output, no tool execution
gemma4:26b	Dense	17 GB	16k	Working — completed multi-step agentic tasks
gemma4:26b	Dense	17 GB	24k	Working but under memory pressure
qwen3-coder:30b	MoE (3.3B active)	18 GB	64k	Unusably slow — too much swap
qwen3-coder:30b	MoE (3.3B active)	18 GB	32k	Still too slow
devstral-small-2:24b	Dense	15 GB	32k	21 GB loaded, tight on memory
gemma4:26b	Dense	17 GB	16k	Crashed after 45 min — infinite repetition loop

Problem 1: Models Can't Find the Tools

Out of the box, both gemma4 and qwen2.5-coder failed to call tools correctly.

**gemma4:26b** tried to call a tool named `shell`, but OpenCode exposes the tool as `bash`:

invalid [tool=shell, error=Model tried to call unavailable tool 'shell'.
Available tools: invalid, bash, read, glob, grep, edit, write, task, webfetch, todowrite, skill, question.]

**qwen2.5-coder:14b** didn't even attempt a tool call — it printed raw JSON as t