Same here. Every time a new local model comes out, I give it a spin with a pretty vanilla coding task ("refactor this method to take two parameters instead of one", or "fix this class of compiler warning across the ~20 file codebase") and more often than not, they get in endless loops, or fail in very unusual ways. They don't yet even approach the usefulness of SOTA models. It's obviously not a fair comparison, though. My 20GB GPU is never going to beat whatever enormous backend Google or Anthropic have.

You can do this with really small models but you have to do a more legwork. I wouldn't expect most trivially small models to handle anything more than 1 file reliably. The new qwen 3.6 is different though, I have heard cases where it is behaving close to sonnet.

That said I don't see why people are so scared to touch code even if it saves them 500 euro a month. Using my IDEs find across my repo and auto replacing 2 patterns is trivial to do and way faster to do by hand. I mostly use small models, it prevents a lot of the issues I've seen with large models and vibe/agentic coding medium to long term. I also write a lot of code.