Never used the cloud models so can't tell about that. My humble experience with llama.cpp + pi agent + Qwen3.6-27B + 3090 24Go VRAM and a codebase of a bit more than 130k is: if you have a workflow where you first draft a PLAN.md then make the model review it, update it with a few iterations adding comments in it like 
<!-- USER: keep this file untouched -->
and implement it Phase by Phase in a git repository it works pretty fine and you can achieve huge amount of work be it refactoring, fixing, adding features... Been doing that for only two weeks when i finally went the agentic way in a sandbox and i'm impressed by what i can do fully local.