---
title: "Ollama Setup Guide for RewriteBar"
description: "Run Ollama models locally on your Mac and connect them to RewriteBar for private AI-powered writing and text transformations."
url: "https://rewritebar.com/docs/ai-providers/ollama"
---

# Ollama

Use Ollama to run AI models locally on your Mac and connect them to RewriteBar.

## Setup Ollama with RewriteBar

1. Install Ollama from [ollama.com/download](https://ollama.com/download).
2. Open Terminal and confirm installation:

```bash
ollama --version

```

1. Download a model (example):

```bash
ollama run llama3.2

```

1. Confirm the model exists:

```bash
ollama list

```

1. In RewriteBar, open **Preferences** → **AI Provider** → **Ollama** and use:  
  - **Base URL**: `http://localhost:11434/v1`
  - **Default Model**: a model from `ollama list` (for example `llama3.2:latest`)
2. Click **Verify and Enable**.

## Model Selection Guide

Start with a smaller model, then move up only if you need better quality.

| Model        | Best For                      | Size    | Notes                                |
| ------------ | ----------------------------- | ------- | ------------------------------------ |
| llama3.2     | Everyday rewriting            | \~2GB   | Best starting point for most users   |
| llama3.1:8b  | Higher quality rewrites       | \~4.7GB | Better output, more RAM needed       |
| mistral:7b   | Balanced quality and speed    | \~4.1GB | Good general-purpose alternative     |
| codellama:7b | Technical and code-heavy text | \~3.8GB | Useful for developer-focused writing |

### How to choose

- Prioritize **speed**: pick `llama3.2`.
- Prioritize **quality**: try `llama3.1:8b`.
- Working on **technical docs**: test `codellama:7b`.
- On limited RAM: stick to smaller models first.

You can switch models anytime in RewriteBar settings without changing your workflow.

## Troubleshooting

- **Connection refused**: Ollama is not running. Start Ollama and retry.
- **Model not found**: run `ollama list` and select a model that is installed.
- **Slow responses**: choose a smaller model or close memory-heavy apps.