Docs · Getting started
Requirements and platforms
Nodeau runs on Linux (Ubuntu 24.04, x86_64, NVIDIA) and on Apple Silicon Macs. Here's what each platform does, what it needs, and the evidence behind every level.
Where Nodeau runs#
Two platforms, each set up in the way that suits it:
- A Linux machine with an NVIDIA GPU. Ubuntu 24.04 LTS on x86_64, with a working NVIDIA driver. Every capability runs here, from one machine to a fleet of them.
- An Apple Silicon Mac. Any M-series Mac. It runs standalone and serves chat, embedding, reranking and vision models on the Mac's own GPU through Metal, several at once.
Nodeau has been run end to end on an RTX 5070 Ti, an RTX 3080, an RTX 5060 Ti and an RTX 2080 under Ubuntu 24.04, and on M3 Pro Macs under macOS 26.5 and macOS 27.
Support levels#
A support level is a claim about evidence, not about intent. There are four, and we use them literally.
| Level | What it means |
|---|---|
| Recommended | What Nodeau is developed and measured on every day. A problem here is a bug that blocks a release. |
| Qualified | Run end to end on named, real hardware, from the exact files a customer downloads, with its scope written down. |
| Experimental | Implemented, with its hardware qualification still to come. Nothing sits here today. |
| Unsupported | Nodeau makes no support claim, and stops at the start with a clear message rather than half-working. |
Platforms#
| Platform | Level | What that rests on |
|---|---|---|
| Linux with NVIDIA, Ubuntu 24.04 LTS, x86_64 | Recommended | Several machines and four NVIDIA cards, through reboots, a power cut and a GPU moving between machines. |
| Apple Silicon, M-series Mac | Qualified | M3 Pro Macs installed from the public path and serving chat, embeddings, reranking and vision on Metal, several models at once. A Mac runs standalone, and the scope is in the table below. |
| Windows through WSL | Unsupported | The GPU path through WSL works differently. install.sh recognises WSL and stops before changing anything. WSL is on the roadmap as planned work. |
| Native Windows | Unsupported | Nodeau publishes builds for Linux and macOS. |
| Intel Macs | Unsupported | Metal execution needs an M-series chip, and install.sh names that as the reason. |
What runs where#
| Capability | Linux with NVIDIA | Apple Silicon |
|---|---|---|
| Local inference with an OpenAI-compatible endpoint | Yes | Yes, through Metal |
| Chat | Yes, with streaming | Yes, with streaming |
| The curated catalog, verified downloads and the model cache | Yes | Yes |
| The fit check before anything starts | Yes | Yes, against unified memory |
run, ps, stop, logs and doctor | Yes | Yes |
| Accounts, plans and offline entitlements | Yes | Yes |
| Embeddings and reranking | Yes | Yes |
| Image input (vision) | Yes | Yes |
| Several models serving at once | Yes, one per GPU | Yes, each on its own port, as unified memory allows |
| Structured output and tool calling | Yes | On your Linux machines |
| Batch inference | Yes | On your Linux machines |
| Bring your own model | Yes | On your Linux machines |
| Several GPUs in one machine | Yes | A Mac has one integrated GPU |
| One model across several GPUs | Yes, NVIDIA cards in one machine | A Mac has one integrated GPU |
| A fleet of several machines | Yes | A Mac runs standalone |
| Fleet upgrades from one place | Yes, with Home Pro and Business | Updated on the Mac itself with nodeau update |
| Scheduling modes, power limits, health, and limits and policies | Yes | Part of a Linux fleet |
On Apple Silicon the GPU shares one pool of memory with everything else on the Mac, so whether a model fits depends on what else is running. See unified memory.
GPUs#
| GPU | Level | Notes |
|---|---|---|
| NVIDIA, on Linux | Recommended | Nodeau enforces no minimum compute capability. On a card it has no measurement for, it works out a careful estimate and says so on the decision. |
| Apple Silicon integrated GPU | Qualified | Through Metal, on the Mac's native execution plane. |
| AMD GPUs | Unsupported | A different runtime and a different memory story, so it's more than a configuration flag. It's on the roadmap under exploring. |
| Intel GPUs | Unsupported | The same. |
Nodeau ships measured profiles for the cards it has run on, and computes a figure for everything else from the model's own architecture and file size, plus a further margin. An estimate is always labelled as one. See measured and estimated.
How much GPU memory you need#
The curated catalog sits on a hardware ladder, and every model says which rung it's on:
| Rung | What it runs |
|---|---|
| 8 GB | A compact general model, an embedding model, a reranker and a compact multimodal model |
| 12 GB | A mainstream general model and a mainstream multimodal model |
| 16 GB | A larger, efficient mixture-of-experts model |
| 24 GB, or several smaller cards in one machine | The flagship general model |
A rung is guidance. The decision is always made by the fit check, on your card, at the moment you ask, and a model can fit a smaller card than its label suggests. See the model catalog.
The NVIDIA driver#
On Linux, have a working NVIDIA driver in place before you install Nodeau.
nvidia-smiIf that prints a table of GPUs, you're ready.
The driver is yours. Nodeau checks it and reports what it finds, and it leaves the driver, Secure Boot, the bootloader and the kernel command line exactly as they are, because those decide whether your machine boots to a desktop. If the driver isn't working, Nodeau stops and tells you what it saw.
Disk#
| The Nodeau binaries | under 100 MB |
| Container images (Linux) | about 3 GB |
| The starter model | about 2.6 GB |
| Recommended free space | 20 GB |
Model files are large, and you'll probably want more than one. Plan for the models you intend to run rather than for the starter.
Memory#
On Linux, 16 GB of system RAM or more is comfortable. Model files pass through system memory on their way to the card.
On Apple Silicon, memory is unified: one pool serves the system, your applications and the GPU. Whether a model fits therefore depends on what else is running, and it can change from one attempt to the next. If a model that started yesterday doesn't fit today, Nodeau is reading the machine correctly: close something and try again. See Install on a Mac.
Network#
Installing needs outbound HTTPS to reach the release, the container images and the model publisher. After that, inference needs no network at all.
The local endpoint binds 127.0.0.1 and nothing else, and that isn't a setting. Nodeau opens no inbound ports and needs no port forwarding. When you connect a fleet to your account, your machines make the connection outward to Nodeau Cloud. See security and privacy.
Privileges#
install.sh | Runs as you, installs into a directory you own, and calls no sudo. It stops if you run it as root. |
nodeau install on Linux | Uses sudo for the package and cluster steps. Every privileged command is printed in full before you're asked. |
nodeau install on a Mac | Needs no password. Everything lives in your own Library folder. |
What each plan allows#
Limits are enforced by the binary, from a signed entitlement it checks locally.
| Home (free) | Home Pro | Business | |
|---|---|---|---|
| Machines | 1 | 3 | Unlimited |
| GPUs in any one machine | 1 | 2 | Unlimited |
| Workloads running at once | 2 | Unlimited | Unlimited |
| Batch inference | From Home Pro | Unlimited jobs | Unlimited jobs |
Seeing your machines at app.nodeau.ai | Yes | Yes | Yes |
Operating them from app.nodeau.ai | From Home Pro | Yes | Yes |
Each GPU runs one workload at a time, so the GPUs you have also set how many workloads run at once.
Everything that keeps you safe, meaning the fit check, artifact verification, authentication and every integrity check, is the same on every plan. See accounts and plans.
These docs describe the current release, the beta channel. Spotted something wrong or missing? Tell us. A report from a machine we have never seen is the most useful thing we get.