The best local LLM for coding, by how much memory you have

A local model won't replace a frontier cloud model for long, autonomous coding sessions. It's very good at the everyday part: explaining code, writing functions and tests, reviewing a diff, and answering "how do I do X in Y". And your code never leaves your machine.

Picks by memory

Memory	Pick	Command	Good for
8 GB	Qwen3 8B	`ollama run qwen3:8b`	Explaining code, small functions, shell one-liners
12 GB	Qwen3 14B	`ollama run qwen3:14b`	Real help with functions and reviews
16 GB	gpt-oss 20B	`ollama run gpt-oss:20b`	Reasoning through bugs, tool use
24 GB	Qwen3 32B	`ollama run qwen3:32b`	The strongest single-card coder here
64 GB+	gpt-oss 120B	`ollama run gpt-oss:120b`	Close to cloud quality on hard problems

Dedicated coding models are released often, and many are worth trying at the same sizes. These general models are a reliable starting point.

Speed matters more for code

Code answers are long, and you often regenerate. Below about 20 tok/s it starts to feel slow. That's why on a 24 GB card many people keep two models:

Qwen3 30B-A3B (qwen3:30b) for fast back-and-forth, at around 160 tok/s on a 3090.
Qwen3 32B for the hard ones, at around 26 tok/s.

Give it enough context

Code eats context. A few files plus the conversation is easily 16,000 tokens, and Ollama's default is shorter, so the model quietly forgets the start. Raise it for coding:

OLLAMA_CONTEXT_LENGTH=16384 ollama serve

That's for Linux, where Ollama runs as a server you start. On Mac and Windows, set the context length in the Ollama app's settings instead. Longer context needs more memory; see how much VRAM you need.

Use thinking mode for bugs

Qwen3 and gpt-oss can reason before answering. It's slower but catches more on real bugs. With Qwen3 add /think to your message; gpt-oss lets you set a reasoning effort of low, medium or high.

Wire it into your editor

Most coding tools can talk to Ollama, because it serves an OpenAI-compatible API at http://localhost:11434/v1. In VS Code, extensions like Continue and Cline let you pick an Ollama model for chat and edits. Point them at that address and pick your model.

Agent-style tools that edit many files on their own work best with the bigger models. On 8 to 14B models, keep them to one file at a time.

What to expect

Good: explaining unfamiliar code, writing tests, regex and SQL, converting between languages, reviewing a diff.
Mixed: large refactors, unfamiliar libraries released after the model was trained.
Weak: long autonomous sessions across a big codebase. That's where cloud models still earn their price.

Not sure what your machine runs? Check every GPU or the homepage.