I Ran AI Models on My Own Machine for a Month: Honest Notes
Local models are genuinely good now and genuinely different. A month of daily use, with the honest ledger of what worked and what did not.
For one month I moved every task that could move onto local models running on my own machine: drafting, summarizing, code completion, and a lot of experimentation. The honest summary: local is real, local is different, and the gap to cloud models is narrower than the benchmarks suggest and wider than the enthusiasts admit.
What genuinely worked: the always-on jobs. Summarization, drafting first passes, rewriting, and code completion ran locally all month with no perceptive quality loss for my purposes, and the privacy relief is not hypothetical; I stopped redacting notes before summarizing them, which is a behavior change that only happens when the data truly does not leave.
What surprised me: latency is the feature nobody prices. A local model answering in half a second changes how you use it the way a fast search box did. I asked it throwaway questions I would never bother a cloud model with, and the throwaway questions turned out to be where a lot of the value lived.
What did not work: the big-reasoning jobs. Long-document analysis, complex multi-step code, and anything where being wrong is expensive still went to the frontier cloud models by week two. Local models fail confidently on exactly the tasks where you most need a second opinion, which is a dangerous combination.
The month ended with a split I expect to keep: local for the high-volume, low-stakes, privacy-sensitive layer; cloud for the deep-work layer. The hardware cost is real and sunk now. The honest framing for anyone considering it: local is not a replacement, it is a second tier, and it is finally good enough to be worth a tier.