What matters for a coding model
- Context window: a bigger context lets the model see more of your file or codebase at once. Coding-specific models often ship with larger context windows than general chat models of the same size.
- Language and framework coverage: training data varies. A model strong in Python may be noticeably weaker in a less common language; check community feedback for your specific stack if you have one.
- Instruction-following on edits: coding work is mostly editing existing code, not writing from scratch. A model that reliably makes only the change you asked for, without rewriting unrelated code, matters more day to day than raw benchmark scores.
Picks by hardware tier
| Tier | Model | Notes |
|---|---|---|
| 8GB | Qwen2.5-Coder 7B | Coding-focused variant of Qwen at a size that fits entry hardware |
| 8GB | Llama 3.1 8B | Not coding-specific, but a reasonable fallback if the Qwen coder variant isn't available for your tool |
| 12-16GB | Qwen3 14B | Community favorite for coding at this tier, not a dedicated coder variant but strong on code tasks generally |
| 12-16GB | DeepSeek-R1-Distill 14B | Better fit for logic-heavy, algorithm-style problems than straightforward CRUD code |
| 24GB | Qwen3-Coder 32B / Qwen2.5-Coder 32B | The strongest coding-dedicated pick most people can realistically run locally |
| 24GB | Devstral | Apache 2.0 licensed, purpose-built for agentic coding workflows rather than single-shot completion |
Run any of these through the hardware calculator with your actual VRAM and preferred context length before downloading; a coding model you'll actually use with a large context window needs more headroom than the same model at a short chat-length context.
Agentic coding vs. autocomplete-style use
There's a real difference between a model that suggests the next few lines as you type, and a model built to run as an agent: reading multiple files, deciding what to change, running commands, and iterating on the result with minimal supervision. Devstral specifically targets the agentic case. A general coding model can often do agentic-style work with the right tooling around it, but a model trained for it tends to follow multi-step instructions more reliably without wandering off task.
If your workflow is mostly inline suggestions in an editor, the tier picks above in a straightforward chat or completion setup are enough. If you want a model driving a multi-step coding task on its own, look specifically for agentic-coding-focused models like Devstral, and expect to need more context headroom for the back-and-forth involved.