Ollama + Aider Local Deployment Complete Guide

This guide covers how to set up a fully local AI coding environment using Ollama (local model runtime) and Aider (AI coding assistant). The advantages of this approach are:

• Completely free: No API costs — configure once, use forever
• Data security: Your code never leaves your machine, ideal for sensitive projects
• No VPN required: Does not depend on any external services
• Offline capable: Once models are downloaded, you can work without an internet connection

Prerequisites

• macOS, Linux, or Windows (WSL2)
• At least 8GB RAM (16GB+ recommended)
• At least 10GB of disk space (for storing models)
• A GPU helps but is not required — CPU works too (just slower)

---

Step 1: Install Ollama

macOS / Linux

curl -fsSL https://ollama.ai/install.sh | sh

Windows

Visit ollama.ai to download the installer, or run the Linux version through WSL2.

Verify Installation

ollama --version
# Output should be something like: ollama version is 0.5.x

---

Step 2: Download a Coding Model

Choose a model that matches your hardware:

Recommended Models

# Lightweight · 8GB RAM sufficient · Recommended for beginners
ollama pull qwen2.5-coder:7b

# Mid-range · 16GB RAM · Stronger coding capabilities
ollama pull qwen2.5-coder:14b

# High-end · 32GB+ RAM · Best local coding model
ollama pull qwen2.5-coder:32b

# DeepSeek Coder (alternative)
ollama pull deepseek-coder-v2:16b

Download Time Estimates

Model	Size	Download Time (100 Mbps)
qwen2.5-coder:7b	~4.7GB	~7 minutes
qwen2.5-coder:14b	~9GB	~12 minutes
qwen2.5-coder:32b	~20GB	~30 minutes

Test the Model

ollama run qwen2.5-coder:7b
# Type: Write a Python quicksort function
# Press Ctrl+D to exit

---

Step 3: Install Aider

Option 1: Using pip (Recommended)

# Using a virtual environment is recommended
python -m venv ~/.aider-env
source ~/.aider-env/bin/activate  # Windows: ~/.aider-env/Scripts/activate

pip install aider-chat

Option 2: Using pipx (Global Install)

pip install pipx
pipx install aider-chat

Verify Installation

aider --version
# Output: aider x.x.x

---

Step 4: Configure Aider to Use Ollama

Aider connects to Ollama through its OpenAI-compatible interface:

# Option 1: Command-line arguments
aider --model ollama/qwen2.5-coder:14b \
      --openai-api-base http://localhost:11434/v1 \
      --openai-api-key ollama

# Option 2: Environment variables (recommended)
export OLLAMA_API_BASE=http://localhost:11434
aider --model ollama/qwen2.5-coder:14b

Create a Configuration File (Recommended)

Create a global config at ~/.aider.conf.yml:

# ~/.aider.conf.yml
model: ollama/qwen2.5-coder:14b
openai-api-base: http://localhost:11434/v1
openai-api-key: ollama
auto-commits: true
dark-mode: true

After this, simply running aider will use the configuration above.

---

Step 5: Use Aider in Your Projects

Basic Usage

# Navigate to your project directory
cd ~/my-project

# Start Aider (automatically loads .aider.conf.yml)
aider

# Or specify files to edit
aider src/main.py src/utils.py

Common Commands

Within the Aider interactive interface:

# Add a file to the context
/add src/components/Button.tsx

# List files currently in context
/ls

# Undo the last change
/undo

# View the git diff
/diff

# Exit Aider
/exit

Practical Usage Example

You: Add pagination to src/api/users.py, following the implementation in src/api/posts.py

Aider: Let me analyze both files...
[Reads users.py and posts.py]
[Generates modification plan]
[Shows diff]

Apply this change? (y/n): y

You: Also add unit tests for the paginate function

Aider: Sure, I'll add tests in tests/test_api.py...

---

Performance Optimization

GPU Acceleration (NVIDIA)

Ollama automatically detects CUDA. If you have an NVIDIA GPU, it will be used automatically:

# Verify GPU is being used
ollama run qwen2.5-coder:7b
# While running, use nvidia-smi to check GPU utilization

GPU Acceleration (Apple Silicon)

Apple M-series chips have native support — Ollama automatically uses Metal acceleration:

# Recommended setup for M1/M2/M3/M4 Macs
ollama pull qwen2.5-coder:14b  # 16GB unified memory can run 14B

Adjust Context Length

# Increase the context window (requires more memory)
ollama run qwen2.5-coder:14b --num-ctx 8192

---

Local vs. Cloud Solutions Comparison

Dimension	Ollama + Aider	Cline + Volcengine Ark	Claude Code
Monthly cost	Free	9.9+ CNY	$20+
Data security	★★★★★	★★★☆☆	★★★☆☆
Model quality	★★★☆☆	★★★★☆	★★★★★
Response speed	Hardware-dependent	Fast	Fast
Offline use	Yes	No	No

---

FAQ

Q: Ollama service isn’t starting

# Manually start the Ollama service
ollama serve

# Or check the service status
curl http://localhost:11434/api/tags

Q: Model responses are very slow

• Running on CPU is expected to be slow — the 7B model generates roughly 3–10 tokens/second on CPU
• Consider using a smaller model (3B/4B)
• If you have a GPU, verify that your drivers and CUDA version are compatible

Q: Aider says it can’t find the model

Make sure the Ollama service is running and the model has been downloaded:

ollama list  # View downloaded models
ollama pull qwen2.5-coder:7b  # Re-download if needed

Q: Can I use a more powerful model?

If you want stronger models without fully relying on the cloud, consider:

• Hybrid approach: Use Ollama locally for simple tasks, and connect Aider to the DeepSeek API for complex tasks (~$0.07 per million tokens)
• Configuration: Aider supports configuring multiple models simultaneously — use the /model command to switch between them

---

Advanced Configuration

Project-Level Configuration

Create .aider.conf.yml in your project root:

# .aider.conf.yml (project-level, takes priority over global config)
model: ollama/qwen2.5-coder:14b
auto-commits: false  # Manually control commits in this project

Integration with Git Hooks

# Add to .git/hooks/pre-commit
# Have Aider check code quality before each commit

---

You’ve finished setting up your local AI coding environment! While the model capabilities don’t match top-tier cloud services, this setup is more than sufficient for everyday tasks like code completion, function generation, and refactoring — and it’s completely free with full data privacy.

If you have any questions, feel free to discuss on GitHub Issues.