I don’t need the largest available model to propose filenames. I might need it for a difficult explanation or a long debugging session. Treating both jobs the same makes local-versus-cloud discussions less useful than they could be.
A local model is worth considering for repetitive work on private files: extraction, classification or a first pass over notes. Whether it is good enough depends on the actual files and the machine. Small doesn’t automatically mean fast, and running locally doesn’t automatically make the surrounding application private.
Check the whole path. Are embeddings generated remotely? Does telemetry include excerpts? Can a failed local request silently fall back to an API? Keeping the model on the laptop means little if the content leaves through another component.
For a hybrid setup, I’d make the remote step explicit. Show what will be sent and let the person decline. Test the weaker local path rather than assuming that an answer of any quality counts as a working fallback.
I like having the option to do useful work without a network connection. I also like good remote models. The interesting decision is which job belongs where, with the data boundary clearly stated.