Over the past few months, I've experimented with several workflows for using LLMs in software development. In this article, I describe the local setup I've validated on my own PC: Ollama running on Windows, Aider inside Ubuntu on WSL2, and a code-focused model running entirely on the local machine.
The machine I used for my first experiment runs Windows 11 with an Intel Core i7-1355U CPU, 48 GiB of RAM, and no dedicated NVIDIA GPU. It's generally a capable machine, but not particularly well suited for AI inference. Because it lacks a dedicated GPU, I had to run the model in CPU-only mode.
The goal of this setup was to create an environment where both the source code and the data remain entirely on the local machine—far from the cloud-based frontier model approach.
From an architectural standpoint, I decided to keep Ollama on Windows because I also use the installed models with other Windows applications, and I wanted to avoid maintaining duplicate model installations. Aider, on the other hand, runs inside WSL, where installation and configuration are simpler and the Linux command-line environment is more convenient.
Ollama remains exposed on the local port 11434. Aider, running inside Ubuntu WSL2, connects to Ollama through the OLLAMA_API_BASE
environment variable.
This approach avoids duplicating models inside WSL while keeping development repositories in the Linux filesystem under ~/repos
, where Git and other command-line tools work much more naturally.
Since this machine is not an AI workstation, the solution had to be realistic. That meant avoiding massive GPU-hosted models, cloud APIs, and overly complex automation.
The final architecture looks like this:
Windows
├─ Ollama
│ └─ qwen2.5-coder:14b
│
└─ WSL Ubuntu
├─ Aider
├─ Git
└─ Development repositories
Keeping Ollama on Windows while running Aider inside WSL was a natural choice. I already use Ollama with several Windows applications, so duplicating the models inside WSL made little sense. At the same time, I much prefer working with the Linux command line for software development.
I installed Ollama on Windows and downloaded the primary coding model:
ollama pull qwen2.5-coder:14b
I also installed the smaller model for quicker testing:
ollama pull qwen2.5-coder:7b
To verify the installation:
ollama list
Initially, Aider running inside WSL couldn't communicate with Ollama running on Windows.
To make it work, I had to configure WSL to use mirrored networking.
From PowerShell:
notepad $env:USERPROFILE\.wslconfig
Then I added the following configuration:
[wsl2]
networkingMode=mirrored
dnsTunneling=true
firewall=true
autoProxy=true
Finally, I restarted WSL:
wsl --shutdown
After reopening Ubuntu, I tested the connection:
curl http://127.0.0.1:11434/api/tags
The response confirmed that WSL could successfully reach the Ollama server running on Windows at 127.0.0.1:11434
.
Inside Ubuntu/WSL:
sudo apt update
sudo apt install -y git python3 python3-pip python3-venv pipx curl
python3 -m pipx ensurepath
exec $SHELL -l
pipx install aider-chat
To verify the installation:
aider --version
I also configured Git so I could version-control my experiments:
git config --global user.name "xxxx"
git config --global user.email "xxxxxxx@xxxxx.xx"
I created a simple test project:
mkdir -p ~/repos
cd ~/repos
mkdir my-project
cd my-project
git init
Then I launched Aider:
cd ~/repos/my-project
OLLAMA_API_BASE=http://127.0.0.1:11434 aider --model ollama_chat/qwen2.5-coder:14b --no-show-model-warnings
When Aider displayed:
Aider v0.86.2
Model: ollama_chat/qwen2.5-coder:14b
Git repo: .git
I finally knew the basic setup was working correctly.
For convenience, I also wanted to access the repository directly from Windows Explorer.
From inside WSL:
explorer.exe .
This command opens the current WSL directory directly in Windows File Explorer.
Once I had Aider working with a local Ollama instance, there was still one thing I wanted to understand better: what was happening under the hood. How many tokens were being processed? How long did inference take? What kind of throughput could I expect?
We'll cover all of that in the next installment. 😉