Code-specific models
Models trained specifically on code tend to punch above their size for programming. Qwen2.5-Coder (from Alibaba's Qwen team) is a strong recent pick, comes in several sizes from small enough for a laptop up to larger GPU builds, and has a long context window for working across files. DeepSeek-Coder is another well-regarded code-focused family with openly licensed weights. Code Llama, Meta's code-tuned Llama variant, is older but widely supported and a safe baseline.
These are the models to reach for when the work is autocomplete, refactors, writing tests, or explaining unfamiliar code — the high-volume tasks where a local model removes the per-token meter entirely.
General models that also code
You don't always need a code-specific model. General open models like Meta's Llama 3 are competent programmers and double as everything-else assistants, which is handy if you'd rather keep one model loaded. They trade a little code specialism for broader range.
The honest note: none of these match a frontier cloud model on the very hardest problems. They're chosen for being good enough on the everyday majority while running free and private on your own machine.
How oi fits
oi is model-agnostic by design. You pull whichever model above suits your hardware through Ollama or llama.cpp, point oi at the runtime, and use it from the CLI and the VS Code extension. Switching models is changing one setting, so you can try a few and keep the one that feels best on your machine.
Open coding models worth running locally
| Model | Typical sizes | Strength |
|---|---|---|
| Qwen2.5-Coder | From laptop-size up to large GPU builds | Strong code model, long context |
| DeepSeek-Coder | Small to large variants | Code-focused, openly (MIT) licensed weights |
| Code Llama | 7B / 13B / 34B | Widely supported code baseline |
| Llama 3 | 8B / 70B | Capable general model that also codes |
frequently asked
Which model should I start with?
A code-specific model around 7B, like a small Qwen2.5-Coder or DeepSeek-Coder build, is the easiest first step — it runs on most modern machines and is tuned for programming. Move up in size if your hardware allows and you want more capability.
Are these models actually free?
The weights are open and free to download. DeepSeek-Coder, for example, is MIT-licensed. Always check the specific license for your use, but for personal and most commercial coding use these are free to run.
Do code-specific models beat general ones?
For pure coding tasks, a code-tuned model often does more with less size. A general model like Llama 3 is the better pick if you want one model for both coding and everything else.
Can oi run all of these?
Yes. oi is model-agnostic — it runs whatever you've pulled into Ollama or llama.cpp, so any of these works and you can switch between them.
Last updated June 19, 2026