Every prompt you type into a cloud AI is a confession. It travels to a server you do not own, gets logged against an account tied to your identity, and joins a permanent record of what you wanted, feared, planned, and asked at 2 a.m. when you thought no one was watching. That record is worth more than the subscription you pay. You are not the customer. You are the corpus.
Now the plot twist that the platforms would rather you not dwell on: you can run a capable AI entirely on your own laptop, offline, where not one byte leaves the machine. No account. No log. No token meter ticking. The same class of model that felt like magic two years ago now fits in a file you download and a program you launch with one command. The cloud was never a technical necessity. It was a business decision.
The order of magnitude
Let us keep this honest with real numbers. A model like Mistral 7B or Llama 3 8B — around seven to eight billion parameters — quantized down to four bits, occupies roughly four to five gigabytes on disk and runs on a laptop with 16 GB of RAM. No datacenter GPU required. On an Apple Silicon Mac or a machine with a modern discrete GPU, it responds at conversational speed. Bigger models — 30B, 70B — want more memory and patience, but the 7-8B tier is already genuinely useful: it drafts, summarizes, codes, translates, reasons through everyday problems.
You will not get GPT-4-class frontier performance on a laptop. Be honest about that. What you get is last year's frontier, running privately, for free, forever. For the overwhelming majority of real tasks, that is not a downgrade. That is liberation.
How to actually do it
This is not a research project anymore. The tooling is embarrassingly simple.
- Ollama is the shortest path. Install it, then type `ollama run mistral` in a terminal. It downloads the weights and drops you into a chat. That is the entire ritual. It runs on Mac, Linux, and Windows.
- LM Studio gives you a graphical app if terminals make you nervous. Browse models, click download, chat. It handles the plumbing.
- llama.cpp is the engine under the hood if you want to build your own tools on top — a lean C++ runtime that made all of this possible on consumer hardware.
Once it is running, pull the network cable or turn off Wi-Fi and keep talking to it. It does not care. It has everything it needs already on your disk. That moment — watching an AI answer with the internet fully disconnected — is when the illusion breaks. You realize you never needed them to hold the mind for you.
What you get back
The obvious win is privacy, but the shape of it is worth naming precisely.
Your prompts are your inner speech made external. Your medical worries, your legal questions, your business plans, your unfinished thoughts. Every one you send to a cloud model becomes their asset — training data, behavioral profile, a line in a dossier. A local model gives you a thinking partner that forgets nothing you tell it to remember and remembers nothing you do not. The conversation dies with the session unless you choose to keep it. That is the natural state of a private thought.
There is a sovereignty here beyond privacy. A local model cannot be rate-limited when you need it most. It cannot be deprecated, price-hiked, or have its personality "updated" overnight into something you no longer trust. It cannot refuse your legitimate use case because a policy team in another country decided your topic was off-limits. It is a tool that answers to you. In a world of rented cognition, that is rare.
Our record
Weigh the two arrangements on the Scales. On one pan: cognition as a service, metered and logged, your inner life flowing outward to a center that records it. On the other: cognition as a possession, silent, local, yours. The first tilts toward Isfet not because the cloud is evil but because it concentrates — every thought routed through one node, one ledger, one owner. The second restores balance by keeping Ba at home. Your breath-soul does not need to visit a datacenter to think. The local model is a small act of Maat: it puts the mind back where it belongs, inside your own walls.
The honest tradeoffs
I will not sell you a fantasy. Local models have real limits, and pretending otherwise is how people get burned.
- Capability ceiling. The very hardest reasoning, the longest context, the freshest knowledge — the frontier labs still lead there. For those specific tasks, closed models are simply stronger today.
- Setup friction. It is one command, but it is still a command. The first download is a few gigabytes. You need enough RAM.
- No live knowledge. A local model knows only what it was trained on. It cannot browse. Pair it with a local search index if you need current facts.
- Maintenance is yours. No one patches it for you. New models ship constantly; you pull the ones you want.
None of these are dealbreakers. They are the price of ownership, and ownership is cheap compared to the alternative — a lifetime of renting your own thoughts back from the people who logged them.
Start tonight
You do not need permission. You do not need a budget. You need one download and one command.
Install Ollama. Run `ollama run mistral`. Ask it something you would never type into a cloud chatbox — something real, something yours. Watch it answer with the Wi-Fi off. Feel the difference between a mind you rent and a mind you hold.
The cloud wants your thoughts on its ledger. Keep them on your disk. Not your keys, not your coins — and not your model, not your mind.