You don’t need to understand any of this to use Mind: that’s the whole idea. But if you’re curious what your computer is actually doing when it answers you, here’s the honest tour — no mechanism, no mystery, no cloud.
Say you type “what should I name a grey cat?” and press Enter. Here’s everything that happens, in order, and where each part of it happens: which is to say, entirely on your desk.
The model doesn’t read English; it reads tokens — small numbered pieces of text. Your question is chopped into a handful of them in a fraction of a millisecond. This happens in memory, on your machine. The original sentence never travels anywhere.
An AI model is billions of numbers sitting in a file on your drive. Mind’s founding observation is that answering a question means reading those numbers in the right order at the right moment — and reading fast, in order, is exactly what a modern SSD and CPU are built for. So the weights flow from your disk through your processor like water through a mill: no giant graphics card required to hold the whole lake at once.
Modern AMD processors carry vector instructions (AVX2, and AVX‑512 on newer chips) that crunch long rows of numbers in single strokes. Mind was written from scratch to keep those units fed. The “thinking” is arithmetic, and your desktop does arithmetic astonishingly well when the data arrives in the right shape.
The answer isn’t computed all at once; it’s produced a token at a time, each one chosen from everything the model has read plus everything it has said so far. That’s why the reply streams onto your screen like typing: you’re watching the computation live, not a loading animation.
The question, the answer, and the model all live on your disk before, during, and after. There’s no server side, no log in someone else’s building, no account that saw it happen. “Smokey,” by the way. Good grey-cat name. Nobody else knows you asked.
The industry decided years ago that AI needs GPUs, and for training — the making of models — it mostly does. But running a finished model is a different job. Inference doesn’t create anything; it reads. And we treat that reading as what it is: a streaming problem, the kind CPUs and fast local storage have been quietly excellent at for decades.
The other half is size. We don’t train or distill anything: we take a standard open model, compress its number format, and run it faster on our engine. The model’s knowledge is untouched — the same billions of numbers, stored leaner so your machine can move them quicker. What the model knows is its business; how fast it moves is ours.
Put those together and the GPU stops being the price of admission. The machine on your desk clears the bar — and the board shows by how much.
Buy a license and download Mind from your account. One small binary; no framework, no driver stack, no container.
One key ties the license to this machine. From then on, Mind runs with or without an internet connection.
Point Mind at an open model — it walks you through fetching one — and start asking. Every token is computed on your CPU, on your desk.
The rule of thumb is simple: a model needs roughly its own file size in free memory, plus headroom for the conversation. Smaller number formats mean smaller files — the same model at INT4 is about half its INT8 size — so the models most people run daily fit comfortably in an ordinary desktop or laptop.
The processor matters more than its age suggests. Any AMD chip with AVX2 — which covers most machines from the last decade — can run Mind. Chips with AVX‑512 run it considerably faster. When Mind starts, it tells you exactly what your CPU supports; no guesswork, no spec-sheet archaeology.
Not sure where your machine lands? The built-in benchmark answers in about a minute, with a number you can hold up against the board.
Open model families we’ve validated end-to-end: Phi‑3, Llama 3, and the Qwen line from 3B up through the big ones. The supported list ships with Mind and grows with it; if a model is on the list, it’s been benchmarked and regression-tested, not just “known to load.”
It’s the same open model — we don’t retrain or trim it. We compress its number format so it moves faster, a standard technique with a small, measurable footprint. The tradeoff is on the supported list per model, in numbers, not vibes.
No. Prompts, answers, and models live on your disk. Mind touches the network once, at activation, and then works whether or not a network exists. Airplane mode is a perfectly good privacy audit: try it.
Mind is deterministic on your machine: identical question, identical settings, identical answer, token for token, run after run. If you want variety, turn the temperature up; if a result matters, you can reproduce it exactly. Most AI can’t promise that. Ours is built to.
Mind is CPU-first by design — the point is that the machine you already own is enough, no thousand-dollar accelerator required. If you own a monster GPU, enjoy it for what it’s great at; Mind isn’t competing for it.
Not today. Mind is written for x86-64 AMD processors on Windows and Linux, and we’d rather say “not yet” plainly than ship something we haven’t proven. When that changes, it’ll change with benchmarks attached.
A file on your disk, a processor that reads it fast, and an answer that never left the room. The rest is just using it.