A quick field note, because people ask me now and then: which model do I run locally.
Mostly Qwen 3.5 Coder on an M5 Max, and honestly it is already pretty usable. Sure, you cannot compare it to Sonnet or Opus 4.6, but it feels roughly like the Sonnet 3.5 I was using last summer, give or take.
Simpler things it one-shots (a snake game, say 🤣), otherwise you have to herd it more. It is not enough to tell it what you want, you mostly have to tell it how you want it, what to use, and so on. But it works.
I use it mainly for things I do not want to send to an online LLM: typically a production deployment or work with more sensitive data. When the data is not allowed to leave my machine, there is nothing to discuss, the model runs here.
And for something running entirely on local hardware, it is genuinely impressive. Think about where we were even a year ago, the jump is huge. Local models stopped being a hobbyist toy and turned into an actual tool for real work.
For me it fits how I like things in general: under my own control, on my own hardware, without my data flowing through someone else's servers. I still use the cloud for heavy reasoning, but it is good to have the option to reach for something that never leaves the desk.