5 min readVarun

Which local coding model should you run?

The honest short answer: Qwen2.5 Coder, at the largest size that fits your GPU with headroom to spare. Everything below is the reasoning and the exceptions.

Start with what fits

A model that barely fits will spill into system memory and crawl. The rule we use is that you want the model's VRAM figure to be comfortably under your card, not equal to it, because your OS and browser also want the GPU.

4 GB
Qwen2.5 Coder 3B — usable for edits and tests
6-8 GB
Qwen2.5 Coder 7B — the sweet spot for most people
12 GB
Qwen2.5 Coder 14B — noticeably better at refactors
24 GB
Qwen2.5 Coder 32B — closest to a cloud coder

What actually improves with size

Going from 3B to 7B is the biggest single jump: instruction following gets reliable and the agent stops losing the thread mid-task. From 7B to 14B, what improves is multi-file coherence — it holds a refactor together across files instead of fixing one and forgetting the other. Past 14B the gains are real but smaller, and you pay in speed.

When not to use a coder model

Coder models are tuned to emit code, which makes them worse at explaining. For Ask and Analyze — understanding a codebase, writing up an architecture — a general model of the same size does better. Qwen2.5 14B and Gemma 3 12B are both stronger teachers than the coder SKU at the same VRAM.

Apple Silicon

Unified memory counts as usable VRAM, which makes Macs punch above their weight for local inference. A 36 GB machine runs models a 24 GB discrete card cannot. Leave more headroom than you would on a dedicated GPU, since the OS shares the same pool.

A note on the numbers

The VRAM figures on our model pages assume the 4-bit quantisation Ollama pulls by default, and the scores are our own guidance from building with these models rather than benchmark results. We would rather give you a number we stand behind than a leaderboard position that does not survive contact with your codebase.