local coder (Qwen3)
A 30B open-weight coder served on my own machine. Mechanical legs run here for zero dollars before any paid token is spent.
What it is
An open-weight 30B coding model served locally on my Mac, wrapped so agent sessions can call it like any other executor. It's the free bottom rung: inventories, conversions, boilerplate, first-draft single-function edits, rubric checks.
Why
The floor of the ladder is free. Legs that would be a paid model's easiest work run on my own hardware for nothing — and the habit of asking “could the local model do this?” keeps the whole spend honest.
How I use it
Invoked through an isolated CLI profile that skips the heavyweight session config — the model stays warm and answers in seconds. Memory is the real constraint on a laptop: one context-capped model resident at a time, a few calls in flight, stop the server when the batch is done.
Get it
Serve a strong open-weight coder with Ollama or similar, cap its context so it fits your RAM, and route your smallest delegated legs there first. The bar is 'good at mechanical work,' not 'good at everything.'
Connected
This thread is scoped to local coder (Qwen3). Opening another node starts a new one.
Ask this node