Hosted open-model chat, desktop runners, one local stack, offline tasks, hardware, updates, and sharing a machine with family.
- 1 Easiest path: a hosted open model chat UI If you want Llama, Qwen, or DeepSeek-class chat without a driver install, open a hosted UI such as Groq, Together, or Fireworks. Spend fifteen minutes on one named model and one public task, then read retention, because the prompt still leaves.
- 2 Friendly desktop runners from zero Desktop runners like Ollama and LM Studio get open weights onto a laptop. Install one app from the official site, pull one small local model, send one public prompt, then find the cloud toggle and the models folder before a 4 GB file shows up three times.
- 3 One local stack end to end (Ollama-style) Pick one local stack and live in it. Three apps on three ports is a raffle, not a setup. This guide walks a 30-minute Ollama path and shows where the model file and the RAM live.
- 4 First useful offline tasks Use a local model already on disk to rewrite, outline, extract fields, or quiz yourself from your own file. Skip live web facts, citations, current prices, and anything that must be true tomorrow. Pack the model on wifi before you fly.
- 5 Hardware reality: RAM, GPU, heat, battery Apple Silicon shares one memory pool; a PC splits RAM and VRAM. A 7B Q4 on 16 GB is the honest starter. A GPU speeds small models but is not required, and heat plus battery are the tax, not a 4090 for email.
- 6 Updating models without breaking your setup A short model name can point at a new file overnight while your prompts stay the same. Pin a known-good copy, test one canned prompt, then delete the loser so disk space and answer quality both stay put.
- 7 Sharing a local setup with family or a small team A kitchen laptop running a local AI model is a shared brain: same login means same chat history. Separate users, keep the app on localhost, and quit it on cafe Wi-Fi so family privacy and small-team sharing stay intentional.