Questions people ask us.

Software that turns the GPUs and computers you already own into one private AI cloud. It checks each machine, works out which models fit on which GPU, downloads and verifies them, runs them, and gives you an OpenAI-compatible API to call them. Add more machines and they become one fleet you can see and steer from anywhere.

Anyone with a capable GPU who'd rather use it than rent one. That's someone with a gaming PC, a developer with a Mac, a home lab with a few machines, or a team with workstations and servers they'd like to run as shared AI infrastructure.

The three plans follow that path: Home for one machine, Home Pro for a personal fleet of up to three, and Business for an organisation.

Three commands on Linux:

curl -fsSL https://get.nodeau.ai/install.sh | bash
nodeau install
nodeau quickstart

On an Apple Silicon Mac, the third one is nodeau run qwen3.5-4b-q4km. The install page walks through both.

About ten minutes of setup, plus the model download. The starter model is about 2.6 GB, so on a slower connection the download is the long part.

Not at all. On Linux, Nodeau runs a small Kubernetes underneath and handles all of it for you. Day to day you use a handful of commands and an ordinary API. And if you do know Kubernetes, it's your cluster, so you're free to look around.

Nodeau Home is free on one machine with one GPU, with no account and no card. Home Pro is $9.99 a month for up to three machines, and Business is built around your organisation. See the plans.

NVIDIA cards on Linux, several per machine, and the built-in GPU of an Apple Silicon Mac. Nodeau has run end to end on an RTX 5070 Ti, RTX 5060 Ti, RTX 3080 and RTX 2080 under Ubuntu 24.04, and on an M3 Pro Mac.

For a card it hasn't measured yet, Nodeau works out a careful estimate from the model's own shape, keeps extra memory in reserve, and tells you that the number is an estimate. As more cards get measured, those estimates turn into measurements.

Yes! On an Apple Silicon Mac, Nodeau runs chat, embedding, reranking and vision models natively on the Mac's GPU through Metal, and several can serve at once, each on its own port. Installing needs no password and no drivers. A Mac runs standalone, as your own private AI endpoint.

Because a Mac shares one pool of memory between everything, whether a model fits depends on what else you have open. Nodeau checks each time and shows you the numbers. Fleets, batch jobs, structured output, tool calling and your own imported models run on Linux machines with NVIDIA GPUs.

Your driver stays yours. It decides whether your machine boots to a desktop, so Nodeau leaves that choice with you: install your distribution's NVIDIA driver, check that nvidia-smi shows your card, and Nodeau takes it from there. It checks the driver when it starts and tells you exactly what it found.

Yes. Run nodeau fleet invite on the machine you have and nodeau join on the new one, and they're a fleet. Each machine reports what it has and checks its own copy of every model. Home Pro runs up to three machines, and Business as many as you have.

Nodeau sees it as a separate resource and starts placing work on it. Each GPU runs one model at a time, so a second card means a second model running side by side, or a second batch worker. Home Pro puts up to two cards in any one machine to work.

Yes, on Linux, when the model is too big for either card alone. Run it with --gpus 2 and Nodeau splits it across two cards in the same machine, checking that each card can hold its own share.

This adds room rather than speed. For more throughput, run separate copies on separate cards, which is what batch workers do.

Nodeau runs on Linux and Apple Silicon Macs today. Windows through WSL is planned, and AMD GPUs are something we're exploring. The roadmap tracks both.

A curated catalog chosen to cover the cards people actually own: a 2.6 GB starter for 8 GB cards, mainstream chat and vision models for 12 GB, a mixture-of-experts model for 16 GB, and a 27B flagship for 24 GB or two cards in one machine. There's also an embedding model and a reranker. Each is pinned to an exact file from its publisher.

nodeau model list shows what you can run, and nodeau model status <model> shows which of your GPUs can hold it.

Yes, if it's a GGUF file. nodeau model import reads the file's header to work out what the model needs and hashes the bytes so the file is its own identity. Then nodeau model qualify starts it on a real card and tests each capability, so you know what it can really do before you rely on it. Your weights stay on your machines. This runs on Linux with NVIDIA.

It takes what your GPU can really give a workload (the memory it can address, minus what your desktop and other programs are using, minus a small safety reserve) and compares it with what the model needs at your context size and concurrency, plus a margin.

The model's figure is a measurement from hardware like yours when Nodeau has one, or a careful estimate from the model's architecture otherwise. The decision always says which.

Because it plans with the memory a model can really use. Part of every card sits outside what the compute API can reach, your desktop holds some of it, and a model that fits with nothing to spare fails the moment something else asks for memory. Planning against the usable number is what keeps models running.

Nodeau tells you before it starts anything, with the numbers: what the card can give, what the model needs, and by how much it's short. It also suggests what would change the answer, like a shorter context or a smaller model. Your download is kept either way.

Straight from their original publishers, over HTTPS, pinned to an exact revision. Every download is checked against a known SHA-256 before it's used, and a file with the wrong checksum is set aside where nothing will run it.

Yes. Point the app or SDK at http://127.0.0.1:8080/v1, use the key Nodeau generated (nodeau auth show), and chat completions, streaming, embeddings and model listing work the way you'd expect.

Yes, with Home Pro or Business on Linux. Hand Nodeau a file of chat or embedding requests and it queues the job, waits for a free GPU, runs it, and writes a results file where every answer is matched to its request. Add --workers 2 and two copies of the model share the work, each on its own card.

Both! nodeau dashboard opens a local view of your hardware, models, services and plan, served by Nodeau on your own machine. And once you connect a fleet, your account at app.nodeau.ai shows your machines from any browser.

Very much so. Every command that prints a table also speaks --json, and the exit codes tell a script whether it got a refusal, a timeout or a mistyped command. The CLI reference lists them all.

It runs the same fit check on every GPU in the fleet, then picks by your scheduling mode: Efficiency, Balanced or Performance. Balanced is the default. nodeau placement explain shows the winner, the runners-up, and why.

A running workload stays exactly where Nodeau placed it, even when another GPU would score better. Your service keeps serving. New placements are where Nodeau's choices come in.

Its models stop with it, the way they would on any single computer. Nodeau keeps each service's definition, so there's nothing to recreate, and with Home Pro or Business you can recover a workload onto another machine from your account. Automatic recovery is in progress.

Yes. One fleet, one place that decides where work runs, and your organisation's own limits on what may run, plus usage in accelerator-hours and a record of every change. Separate accounts for each person, with roles and single sign-on, are planned for Business.

Start with nodeau fleet upgrade plan. It shows which machines would change, which are already current, the order they'd go in, and which workloads would restart, before anything moves. Happy with it? nodeau fleet upgrade apply rolls that exact plan out one machine at a time, workers first and the control-plane machine last, checking each machine before the next. If a step needs attention, the rollout holds there until you resume or cancel it. Fleet upgrades come with Home Pro and Business.

Your prompts, answers, embeddings, images and batch records stay on your hardware. The model runs on your GPU and requests reach it over the loopback interface.

If you connect a fleet to your account, it reports which machines you have, what's plugged into them and what's running, so you can see it from a browser. And a third-party app you point at your endpoint can do whatever that app does, so it's worth knowing what yours does.

Only the machine it runs on. The endpoint listens on 127.0.0.1 and every request needs the key Nodeau generated on that machine. To reach a model from elsewhere on a network you control, put your own reverse proxy in front of it. The API guide covers it.

Only to download new models or updates. Once a model is on your machine, inference is entirely local, and your plan is a signed entitlement the machine checks by itself.

Your machines start every connection, outbound over HTTPS, which keeps your network boundary simple. Models keep serving and scheduling keeps working whether or not the cloud is reachable.

Only when you want one. Nodeau Home installs and runs with no sign-in at all. An account adds your machines in a browser, a paid plan when you want one, and your organisation's usage and history.

Home is one machine with one GPU, free. Home Pro is up to three machines with up to two GPUs in each, and adds batch inference and running your fleet from a browser: starting and stopping models, draining machines, reading logs and recovering work.

Business plans are tailored to each deployment: how many machines, which cards, and what support you need. Talk to us and we'll build the right setup together.

Your machines carry on. Plans are signed entitlements checked on each machine, so inference, batch jobs and the local dashboard keep going regardless.

Any time. nodeau uninstall shows its plan and asks first. It removes what Nodeau added and keeps your driver, your downloaded models and anything that was there before Nodeau.

Run nodeau support bundle and email the file it writes to founders@nodeau.ai. It's scrubbed of keys and credentials and holds no prompts or answers. If setup stopped before the CLI worked, send ~/.local/state/nodeau/install.log instead.