# Local Qwen cheat sheet

Companion to the guide "Run AI on Your Laptop for Free".
Checked against ollama.com/library/qwen3.5 and the Rapid-MLX
repository on 2026-09-01.

## 1. Find your installed RAM

- macOS: Apple menu, About This Mac. Read Chip and Memory.
- Windows: Settings, System, About. Read Installed RAM.
- Linux: `free -h`, total column of the Mem row.

Use the total, not the available figure.

## 2. Pick the size, then run it

| Installed RAM | Command                   | Download |
| ------------- | ------------------------- | -------- |
| 8GB           | `ollama run qwen3.5:2b`   | ~2.7GB   |
| 8GB, tight    | `ollama run qwen3.5:0.8b` | ~1.0GB   |
| 16GB          | `ollama run qwen3.5:4b`   | ~3.4GB   |
| 24GB          | `ollama run qwen3.5:9b`   | ~6.6GB   |
| 32GB          | `ollama run qwen3.5:27b`  | ~17GB    |
| 64GB+         | `ollama run qwen3.5:35b`  | ~24GB    |

Apple silicon: add `-mlx` to any tag, e.g. `qwen3.5:4b-mlx`
(~4.0GB for the 4B). Not available on Intel Macs.

Ollama needs macOS Sonoma 14 or newer.
Type `/bye` to leave a chat. The model stays on disk.

If the reply crawls, run the next size down.

## 3. Continue config.yaml

```yaml
name: Local Qwen
version: 0.0.1
schema: v1

models:
  - name: Local Qwen
    provider: ollama
    model: qwen3.5:4b
    roles:
      - chat
      - edit
      - apply

  - name: Qwen Coder autocomplete
    provider: ollama
    model: qwen2.5-coder:1.5b
    roles:
      - autocomplete
```

Pull the autocomplete model first:
`ollama run qwen2.5-coder:1.5b`

If Continue sees no local models, run `ollama serve`.

## 4. Apple silicon: Rapid-MLX

```bash
brew install rapid-mlx
rapid-mlx serve qwen3.5-4b-4bit
rapid-mlx launch continue-dev
```

Its own RAM recommendations:

| RAM        | Model                |
| ---------- | -------------------- |
| 8 to 15GB  | `lfm2.5-2.6b-4bit`   |
| 16 to 17GB | `qwen3.5-4b-4bit`    |
| 18 to 23GB | `qwen3.5-9b-4bit`    |
| 24 to 31GB | `bonsai-27b-2bit`    |
| 32GB+      | `qwen3.8-27b-4bit`   |

Serves on http://localhost:8000. `rapid-mlx launch` also
accepts `claude-code` and `cursor`. Run `rapid-mlx models`
to see what it suggests for your machine.

## 5. Before you paste anything sensitive

Offline privacy holds only while the model and the route
stay local. Ollama publishes `:cloud` tags and Rapid-MLX can
route to a cloud provider. Check the tag you are running.

## Official sources

- https://ollama.com/library/qwen3.5
- https://docs.ollama.com/quickstart
- https://docs.ollama.com/macos
- https://docs.continue.dev/customize/model-providers/top-level/ollama
- https://docs.continue.dev/customize/deep-dives/autocomplete
- https://github.com/raullenchai/Rapid-MLX