Open Source AI Coding Tools

Why Open Source Matters

Open-source AI coding tools provide transparency, customizability, and freedom from vendor lock-in. They're essential for organizations with data sovereignty requirements, offline development needs, or custom model fine-tuning requirements.

Open Source Models

Code-Specific Models

• DeepSeek-Coder V2: 236B MoE model competitive with GPT-4 on coding benchmarks. Available in smaller sizes (1.3B-33B) for local inference.
• Qwen2.5-Coder: Best-in-class for its size range (7B-72B). Excellent on Apple Silicon via MLX.
• CodeLlama: Meta's purpose-built code model. Available in 7B-70B sizes with infilling capability.
• StarCoder 2: BigCode's latest model trained on The Stack v2. Strong at 15B parameters.

Open Source Tools

Model Runners

• Ollama: One-line model installation and serving. The Docker of LLMs.
• llama.cpp: C++ inference engine supporting GGUF quantized models on CPU and GPU.
• MLX: Apple's framework for efficient inference on M-series chips.

IDE Extensions

• Continue: Open-source AI code assistant that works with VS Code and JetBrains. Connects to any model via OpenAI-compatible APIs.
• Tabby: Self-hosted AI coding assistant with VS Code and Vim extensions.
• Aider: Terminal-based pair programming tool that commits changes to git. Works with any model.

Building a Free AI Coding Stack

Combine: Ollama (model serving) + Qwen2.5-Coder-32B (model) + Continue (IDE extension) for a completely free, privacy-preserving AI coding setup. On an M2 Pro Mac, this provides usable code completion and chat at ~30 tokens/second.