LocalLLaMA hunts cheap Mac coder

The hard limit here is the machine, not the ambition: an Intel-era Mac will favor smaller models and lighter tooling over “best” models that look great on benchmarks.

• –A patching workflow matters more than chat polish, because file edits and commit messages are the real day-to-day tasks.
• –On this class of hardware, 7B-ish code models and local runtimes are the practical ceiling; bigger models will be slow enough to kill momentum.
• –Cheap hosted APIs are often the best value if you can live with internet dependence.
• –If privacy or offline use matters, expect a narrower context window and more manual supervision.