# Run bigger AI models on your own machine with Ollama

The Local Model Lab page runs small models in your browser. Ollama is the next
rung: a free desktop app that runs larger models — 8 billion parameters and up —
on an ordinary Windows or Mac computer, still fully local. Nothing you type
leaves your machine.

## What you need

- A Windows 10/11 or Mac (Apple silicon or Intel) computer from roughly the last
  six years.
- At least 8 GB of memory (16 GB is more comfortable for the larger models below).
- About 5–10 GB of free disk space per model.
- An internet connection for the one-time download.

## Step 1: Install Ollama

1. Go to **ollama.com** and click Download for your system.
2. Run the installer. Accept the defaults.
3. On Windows, Ollama runs quietly in the background (look for the icon in the
   system tray). On Mac, it appears in the menu bar.

## Step 2: Download your first model

Open a terminal:

- **Windows:** press the Windows key, type `powershell`, press Enter.
- **Mac:** open the Terminal app (in Applications → Utilities).

Type this and press Enter:

```
ollama pull llama3.2:3b
```

This downloads Meta's Llama 3.2 (3 billion parameters, instruction-tuned) —
about 2 GB, one time. When it finishes, the model lives on your machine.

Want something more capable? These are good next steps, smallest first:

| Command | Model | Size | Best for |
|---|---|---|---|
| `ollama pull llama3.2:3b` | Llama 3.2 3B | ~2 GB | Drafts, rewrites, quick tasks |
| `ollama pull qwen2.5:7b` | Qwen2.5 7B | ~4.7 GB | Longer drafts, structured output |
| `ollama pull llama3.1:8b` | Llama 3.1 8B | ~4.9 GB | General workhorse |
| `ollama pull qwen3:8b` | Qwen3 8B | ~5.2 GB | Newer, strong reasoning |

Start with the smallest. You can always download a bigger one later.

## Step 3: Chat with it

In the same terminal, type:

```
ollama run llama3.2:3b
```

You get a chat prompt. Type your request, press Enter. Type `/bye` to quit.

Try: *"Draft a three-sentence session description for a conference panel on
volunteer burnout. Plain language, no jargon."*

## Step 4: Use it like a tool, not a toy

- **One task per prompt.** "Summarize these bullets" beats "help with the conference."
- **Give it the shape you want.** "Three sentences," "a table with two columns,"
  "under 50 characters" — small models follow formats well.
- **Verify everything factual.** Local models are honest but limited. Anything
  members will see gets a human check against the source.
- **Keep member PII in here, not in cloud chat tools.** This is the whole point:
  the model runs on your machine, so sensitive drafts never travel.

## Troubleshooting

- **"Out of memory" or very slow:** use a smaller model (`llama3.2:1b` is tiny
  and fast) or close other programs.
- **First response is slow:** normal — the model loads into memory once, then
  speeds up.
- **Weird or repetitive answers:** press Ctrl+C, rephrase with more specific
  instructions, or try the next size up.

## Going further

Ollama has a free desktop app with a friendlier chat window (download from
ollama.com), and the `prompt-pack.md` in this kit is written for these models.
When a task outgrows local models, that's the signal for rung three — a paid
subscription — not for pasting member data into a free cloud tool.