Coding is the hardest everyday job for a local model, for three reasons. The prompts are huge, because an agent sends the files it is working on. The model must follow strict formats, so it can call tools and edit files. And an agent makes many calls in a row, so a slow model is a slow afternoon.
What to choose
- A model tagged for coding or agents. The task tags on each model page and the “Best for” lists show which models are built for it (reading the tags).
- The biggest one that runs well. Coding quality climbs with size more than casual chat does. Prefer one that Runs well over one that is slow but a grade higher (ratings).
- An agent or editor integration. The apps page lists several that work with local models, including OpenCode, Aider, Cline, Zed, Goose and Codex CLI. They connect through the local API.
Settings that matter
- Context. Agents usually need 32K tokens or more. That costs memory (KV cache); set it in the app, and consider an 8-bit cache (how).
- Prompt reading speed. Big prompts mean long waits on machines with slow prefill (two speeds).
- Quantization. Code is less forgiving of low quantization than chat. Stay at Q4_K_M or higher when you can (guide).
- Reasoning models. They plan better but spend many tokens thinking (guide).
Expectations
A local model on a single consumer card will not match the largest hosted models on big, multi-file tasks. It will handle explanations, small edits, tests, shell commands and refactors well, and your code never leaves the machine. Many people use a local model for routine work and keep a hosted one for the hardest problems.