---
title: "CLI Coding Agent — Codex CLI"
canonical: "https://build.nvidia.com/spark/cli-coding-agent/codex.md"
---

# Step 1. Confirm your environment

**Description**: Verify the OS version and GPU are visible before installing anything.

```bash
cat /etc/os-release | head -n 2
nvidia-smi
```

Expected output should show Ubuntu 24.04.3 LTS (DGX OS 7.3.1 base) and a detected GPU.

# Step 2. Install or update Ollama

**Description**: Install [Ollama](https://ollama.com/download) or ensure it is recent enough to support [`ollama launch`](https://ollama.com/blog/launch).

```bash
curl -fsSL https://ollama.com/install.sh | sh
ollama --version
```

If Ollama is already installed, just verify the version:

```bash
ollama --version
```

Expected output should show Ollama v0.15 or newer.

# Step 3. Pull Qwen3.6

**Description**: Download the [Qwen3.6](https://ollama.com/library/qwen3.6) model weights to your Spark node.

```bash
ollama pull qwen3.6
```

Optional variants if you want different memory footprints or precision:

```bash
ollama pull qwen3.6:35b-a3b-nvfp4   # NVIDIA FP4 build tuned for Blackwell (~22GB)
ollama pull qwen3.6:35b-a3b-q8_0    # Higher-quality 8-bit quant (~39GB)
ollama pull qwen3.6:35b-a3b-bf16    # Full precision (~71GB)
```

Expected output should show `qwen3.6` in `ollama list`.

# Step 4. Test local inference (optional)

**Description**: Run a quick prompt to confirm the model loads.

```bash
ollama run qwen3.6
```

Try a prompt like:

```text
Write a short README checklist for a Python project.
```

Expected output should show the model responding. When you are done, type `/bye` or press `Ctrl+D` to exit before continuing.

# Step 5. Install and launch Codex CLI with Ollama

**Description**: Install [Codex CLI](https://github.com/openai/codex), then use Ollama's built-in [launch method](https://ollama.com/blog/launch) to start it against your local model. Ollama configures the local-model integration, but the Codex CLI binary must be installed first. No `~/.codex/config.toml` is required.

```bash
npm install -g @openai/codex
codex --version
ollama launch codex --model qwen3.6
```

Expected output should show Codex CLI starting with Ollama as the provider and Qwen3.6 as the model. Qwen3.6 ships with a 256K context window by default, which is well suited to Codex's agentic workflows.

# Step 6. Complete a small coding task

**Description**: Create a tiny repo and let Codex implement a function and tests.

```bash
mkdir -p ~/cli-agent-demo
cd ~/cli-agent-demo

printf 'def add(a, b):\n    """Return the sum of a and b."""\n    pass\n' > math_utils.py
printf 'import math_utils\n\n\ndef test_add():\n    assert math_utils.add(1, 2) == 3\n' > test_math_utils.py
```

If you do not already have pytest installed:

```bash
python3 -m venv .venv
source .venv/bin/activate
python3 -m pip install -U pytest
```

In Codex:

```text
Please implement add() in math_utils.py and make sure the test passes.
```

Run the test:

```bash
python3 -m pytest -q
```

Expected output should show the test passing. When you are done, run `deactivate` to exit the virtual environment.

# Step 7. Cleanup and rollback

**Description**: Remove the model and stop services if you no longer need them.

To stop the service:

```bash
sudo systemctl stop ollama
```

> [!WARNING]
> This will delete the downloaded model files.

```bash
ollama rm qwen3.6
```

# Step 8. Next steps

- Try the `qwen3.6:35b-a3b-nvfp4` or `bf16` variants for different quality/VRAM tradeoffs
- Use Codex CLI on multi-file changes or test-generation tasks
- Explore the full 256K context window on larger codebases