Start with the wait
Running a model locally is satisfying until every small edit takes longer than doing it yourself. I have found the useful question is less about which model is biggest and more about whether I can keep a short, reliable conversation with it.
For code, that means looking at a few relevant files, making a change and checking it. When the task grows, I would rather write down what I have learned and start the next part with a clearer brief.
A setup I can change
I use LM Studio to serve local models to coding tools. Saved presets make it easier to return to settings that worked, rather than rediscovering a context limit or memory problem every time.
Some jobs fit a quick local model. Others need more time or a hosted model. I try to make that choice around the job in front of me, rather than committing everything to one model.
Keep it manageable
One active model is often easier to live with than several competing for memory. The same goes for tools: I connect what a task needs and keep the rest out of the way. A small change I can check is more useful than a large one I cannot follow.