What changed
Three curves crossed: open-weight models kept shrinking while getting smarter, consumer laptops shipped with real NPUs and unified memory, and runtimes like Ollama made local inference a one-line install.
The result: a mid-range laptop now runs models that would have been state-of-the-art cloud offerings not long ago.
What local handles comfortably today
- Question-answering over your own documents with citations (the retrieval does the heavy lifting).
- Summarising, drafting, rewriting — the bread-and-butter of daily work.
- Meeting transcription with on-device Whisper — audio never uploaded.
- Semantic search and embeddings across tens of thousands of files.
Where the cloud still wins — and how to use it safely
Frontier reasoning on hard problems still favours the largest hosted models. The sovereign pattern is not "never cloud" — it is "cloud by choice, per task, with nothing sensitive attached."
ZSearch treats models as interchangeable: run local by default, hand a hard query to a frontier model when you decide it is worth it — without re-uploading your library either way.