Lost access to a hosted coding assistant overnight, how close can a local setup actually get for daily work?

Memory is the wall, but not in the way it gets discussed. The question is not whether a model fits - it is whether the model plus a useful amount of context fits at the same time.

On 24GB you can comfortably hold a capable mid-sized model. What you cannot comfortably hold is that model and enough context to have read four files properly. So the thing that degrades first is exactly what you guessed, and the reason is the one you guessed, which means your instinct here is sound.

Practical shape of it: single-file work, writing a test, explaining a function, fixing an error with a trace: moved across for me completely and I have not gone back. Anything where the answer depends on holding several files in mind at once got noticeably worse, and worse in a bad way: it does not fail, it confidently edits the wrong layer.

So the hybrid split you are guessing at is roughly the correct one, and the deciding question is "how many files does the answer depend on".

I tried this for four months and went back, and I will give you the honest reason rather than a hardware one.

It was not quality. On the single-file work it was fine and I agree with everyone here about that. It was that I had to think about it. Which model is loaded, is the server up, has it fallen back, is this failing because the code is hard or because the context got truncated. Each of those is small; together they took the tool out of the background and put it in the foreground.

The hosted one was worse at some things and I never once thought about its plumbing.

That may say more about me than about local models, and if your reason for moving is the one in your post: not wanting your access to disappear - then it is a cost worth paying, because the alternative is being here again. But budget for it as an ongoing tax on attention, not a one-off setup weekend. That is the part nobody puts in the comparison.

Adding a wrinkle to the memory answer, because I think tooling gets unfairly let off.

The gap is partly context length and partly that the hosted product was doing a lot of retrieval work you never saw. It was choosing what to put in front of the model, which files, which fragments, in what order. When you run locally you inherit that job, and most local setups do it badly or not at all.

Which means two people with identical hardware and the same model get very different results, and the difference is entirely in what the harness feeds it. That is the part people skip when they say local is not good enough, and it is the part you can actually improve without buying anything.

So before concluding the model is the limit, look at what your setup is actually sending. Mine was sending whole files where a good chunk would have done, which wasted the budget on imports.

Useful, and the split everyone converged on is sharper than what I had.

Rewriting the question for myself: it is not local versus hosted, it is that the retrieval layer was the product and I only ever saw the model. That explains why the single-file cases moved over cleanly and the multi-file ones did not, those are exactly the cases where something has to decide what to look at.

Which means the next thing to work on is not the model and not the card. It is what my setup puts in the window. Going to instrument that before spending anything.