Roadmap
What's here, what's next, and what we're exploring.
Every capability on this page carries one of four labels, so it's always clear what you get by installing Nodeau today and what we're still building.
In the current release. Install Nodeau and it's yours.
Being built or tested right now. It arrives in a future release.
Designed for and on the list. The building starts later.
A direction we believe in and are still thinking through.
You won't find dates here. We'd rather ship a capability than guess at a calendar.
Available
All of this is in the current release (), on a Linux machine with an NVIDIA GPU or an Apple Silicon Mac. It's what you get by following the install guide.
Getting started
One-command installer
AvailableDownloads Nodeau, checks its checksum and puts it in a folder you own.
Setup that shows its plan
Availablenodeau install checks the machine, prints every privileged command, and adopts the pieces that are already there.
nodeau doctor
AvailableA read-only check of the machine and the installation, with a remedy for every finding.
nodeau quickstart
AvailableA verified model, a fit check, a running workload and a local endpoint in one command.
Upgrading in place
AvailableRun the installer again. nodeau update tells you when something newer is out, and can download and verify it for you.
Clean uninstall
AvailablePrints its plan first, and keeps your driver, your models and everything Nodeau found already installed.
Apple Silicon Macs
AvailableChat, embeddings, reranking and vision on the Mac's own GPU through Metal, with several models serving at once, each on its own port. No password needed to install. A Mac runs standalone.
Models and tasks
A catalog with a hardware ladder
AvailableEight curated models, from 8 GB cards up to 24 GB or two cards in one machine. Each one says which card size it's for.
Verified downloads
AvailablePinned to an exact file from the publisher and checked against a known SHA-256 before anything runs it.
Fit checks before anything starts
AvailableA measurement where Nodeau has one, a careful estimate otherwise, and the decision says which.
Accepting a fit risk yourself
Available--accept-estimate-risk, for when the only thing in the way is Nodeau's extra margin on unmeasured hardware.
OpenAI-compatible API
AvailableChat and completions with streaming, on 127.0.0.1 with a key made on your machine.
Embeddings and reranking
AvailableVectors for search and RAG, and a rerank route in the Jina and Cohere shape. On Linux with NVIDIA and on a Mac.
Structured output and tool calling
AvailableJSON that follows your schema, and tool calls with arguments your code can parse. On Linux with NVIDIA.
Vision
AvailableDrop in an image and ask a model what it sees. On Linux with NVIDIA and on a Mac.
Bring your own model
AvailableImport a GGUF file, and qualification proves on a real card what it can do. Your weights stay on your machines. On Linux with NVIDIA.
Batch inference
AvailableA file of chat or embedding requests, results matched to every input, and several workers at once, each on its own card. Home Pro and Business, on Linux.
The Playground
AvailableTalk to the models running on your machine from your browser: streamed chat, images, embeddings and reranking, with the machine, the GPU and the speed beside every answer, and models side by side. Your prompts stay on your machine. On Linux and on a Mac.
Hardware and scheduling
Every GPU scheduled on its own
AvailableEach card in a machine is its own resource and runs one workload at a time. Home Pro puts up to two cards in a machine to work.
One model across several GPUs
AvailableFor a model too big for any one card, Nodeau splits it across cards in the same Linux machine, and each card holds its own share. It adds capacity rather than speed: on our test pair, a model that fit on one card ran about 21% slower split across two.
Scheduling you can steer
AvailableEfficiency, Balanced or Performance, for the whole fleet or one machine. Balanced is the default.
Placement you can read
Availablenodeau placement explain shows why a workload is where it is, and what the alternatives would have cost.
Power limits and budgets
AvailableCap what a card may draw, inside the card's own range, and set a power budget placement has to respect.
Workloads that stay put
AvailableA running workload stays exactly where Nodeau placed it, even when another card scores better. If its machine stops reporting, the work waits for it, unless you've set that workload to recover automatically.
Fleet
Two commands to add a machine
Availablenodeau fleet invite on the machine you have, nodeau join on the new one. Home Pro runs up to three machines.
Your fleet in a browser
AvailableEvery machine, card and workload at app.nodeau.ai, on every plan. Your machines connect out to it.
Remote operation
AvailableStart and stop models, set scheduling, drain machines, read logs and recover a workload from a machine that's gone. Home Pro and Business.
Machine health and alerts
AvailableProcessor, memory, storage, network and GPU for every machine, a day of history, and alerts when something needs you. The readings stay on your hardware.
Drain for maintenance
AvailableNew work goes elsewhere and everything already running keeps serving until you're done.
Each machine checks its own models
AvailableEvery machine hashes its own copy of a model before it's offered work that needs it.
Automatic recovery
AvailableChoose, workload by workload, to have Nodeau bring it back up on another of your machines when its machine stops answering, after a delay you set. Its endpoint stays the same. Home Pro and Business, on Linux fleets.
Model copies across machines
AvailableAsk any machine to keep a verified copy of a model. It fetches the files from the model's publisher and checks every one itself, and a workload set to recover keeps one ready on another machine.
Model locality
AvailableWork runs where a verified copy of its model already is, and a copy makes another machine ready for it, so a start or a recovery skips the download.
Organisation
Your own limits
AvailableQuotas on workloads, GPUs and batch workers, plus which models, cards and machines a fleet may use. Lowering a limit leaves running work alone.
Usage in accelerator-hours
AvailableFrom the scheduler's own record of what was reserved, with an hourly rate of your own if you'd like a cost figure.
A record of changes
AvailableWho changed what, including requests that were turned down, kept for a year.
Know what every machine runs
Availablenodeau update compares every machine with the published release. A release channel only moves forward.
Maintenance windows and release channels
AvailableSet when Nodeau may act on its own, in your own time zone, and whether your fleet follows new releases or stays where it is, from Fleet, then Upgrade in your account. Upgrade plans and rollouts follow both.
Entitlements that travel offline
AvailableCarry a signed entitlement to a machine with no internet. It's checked exactly like one fetched over the network.
Home Pro, on sale
Available$9.99 a month, with checkout and self-service plan changes in your account.
Upgrades
Upgrade planning
Availablenodeau fleet upgrade plan pins a release to an exact build and shows what moving each machine would involve, in order and with the work on it, before anything changes.
Fleet upgrades from one place
AvailableApprove the plan you saw with nodeau fleet upgrade apply, or from Fleet, then Upgrade, in your account. Nodeau moves your machines one at a time, workers first and the control-plane machine last, and checks each one on what it reports before the next begins. If a step fails, that machine goes back to exactly what it ran and the rollout holds there so you can take a look. Home Pro and Business.
In progress
Being built or tested now. Each of these arrives in a future release.
Nothing at this moment: everything we've finished is Available above, and the next work comes from the Planned list below.
Planned
Designed for and on the list.
Scheduled and recurring batch
PlannedBatch jobs that run overnight or on a schedule you set.
Alerts that reach you
PlannedAlerts delivered by email or webhook, as well as on the machine.
Machine history in your account
PlannedA machine's recent readings, visible from your account on the web.
Fleet capacity planning
PlannedAnswering "what could this fleet hold?" as well as "where does this workload go?".
Priority for waiting work
PlannedLetting an organisation say which queued work matters most.
Several people in one organisation
PlannedMembers with roles, groups from your identity provider, and keys for scripts that can only ever do less than their owner. For Business.
Single sign-on
PlannedSign in through your organisation's identity provider. For Business.
Windows through WSL
PlannedRunning Nodeau on Windows machines with an NVIDIA GPU through WSL.
More hardware, measured
PlannedMore cards measured rather than estimated, more model profiles, and more Linux and macOS versions. Tell us about your hardware!
Exploring
Directions we believe in and are still thinking through.
Cloud capacity when your fleet is busy
ExploringYour own hardware first, with cloud capacity when your policy allows and your machines are full.
AMD GPUs
ExploringA different runtime and memory model, which deserves its own careful work.
More runtimes
ExploringvLLM and others alongside llama.cpp.
Automatic quantisation
ExploringFitting a model to your card by picking a smaller quantisation for you.
Scale to zero
ExploringFreeing a GPU while a model is idle and loading it again on the next request.
Git-to-deploy
ExploringA repository as the unit of deployment.
Air-gapped installs
ExploringInstalling from your own mirrors and registries. Entitlements already travel offline.
Want a say in the order?
Design partners with real GPU hardware help decide what we build next. We'd love to hear what you'd like to run.