Michael Limberger

Before we install anything

Let's make sure the machine can actually hold the model.

Look at the chip, then the memory

For Apple Silicon Macs, open Terminal.

Show me.

sysctl -n machdep.cpu.brand_string

You should see something like "Apple M4" or "Apple M2 Max". Then check memory:

sysctl -n hw.memsize | awk '{print $1/1024/1024/1024 " GB"}'

You need 16GB minimum. 32GB or more is better.

Local LLMs load entirely into RAM (unified memory on Apple Silicon). A 30B parameter model needs roughly 20GB just for the weights.

Model size	Minimum RAM	Recommended
7B parameters	8GB	16GB
14B parameters	12GB	24GB
30B parameters	24GB	32GB+
70B parameters	48GB	64GB+

The model we recommend, qwen3-coder, is a mixture of experts: about 30B total parameters, but only about 3B active at once. That makes it faster and lighter than a traditional dense 30B model. Mixture of experts means most of the network sleeps, and only a few experts wake for each token.

Disk, Node, and the wire

Show me. Disk:

df -h ~

You need about 30GB free: roughly 20GB for model files, about 5GB for Ollama, and about 5GB as a buffer.

Claude Code's installer requires Node.js. Check with node --version. If it is missing, get it from nodejs.org or run brew install node.

If Ollama is on a separate machine, verify connectivity. Replace the IP with yours:

ping 10.0.0.79
nc -zv 10.0.0.79 11434

The second command should show "Connection succeeded" or similar.

The terminal, said once

You should be comfortable with cd (change directory), ls -la (list files), cat (view a file), export VAR="value" (set an environment variable), and nano (or vim, or emacs) to edit. If those are foreign, spend 30 minutes on a command line tutorial first. We are not covering terminal basics here.

If Ollama is on a separate machine, that machine's firewall needs to allow incoming connections on port 11434. On macOS: System Preferences, Security & Privacy, Firewall, Firewall Options. Make sure Ollama is not blocked. On Linux: sudo ufw allow 11434.

Before we go on

Checklist: Apple Silicon Mac (or Linux with a GPU), 16GB+ RAM (32GB+ preferred), 30GB+ free disk, Node.js installed, network reachability to the AI server if it is remote, and basic terminal comfort.

Everything on one box? That is simpler. Ollama and Claude Code on the same machine means you use localhost instead of an IP. No firewall dance, and no network latency. The downside: your main machine gets hot and slow while generating. A dedicated AI server lets you keep working while the model thinks.

Next: installing and configuring Ollama.