Welcome · Docs · ResonTech

One place for the whole model lifecycle

ResonTech is a platform for training and serving AI models on any set of NVIDIA GPUs: ours, rented, or the ones you already own. You write the Python of your task; deployment, scaling and operations live inside the platform. The SDK and the console stay the same wherever the compute sits.

How it works: any GPUs, one console

GPUs join from wherever they are, and one console manages and schedules them. You describe the task; the platform places it on machines that fit and reports back.

Managed and scheduled from one console. The lines are control, not one job spread across continents.

1
Connect GPUs
Machines in any location join the platform with our software. The console manages and schedules them.
2
Describe the task
Start from a ready template, or tell your AI assistant what you want.
3
ResonTech job
Your training or inference job runs on those machines. Training runs until its rounds are done; an inference endpoint stays up until you stop it.
4
Watch the results
Your model is ready to respond. Live stats per GPU in the dashboard, results in one place, and a ready API for your app.

Assistants read these docs from /llms.txt. Point yours at it and ask.

GPU workers and the central server

Every job runs on two kinds of machines. GPU workers run your model; you choose them. One central server per job coordinates them; the platform provides it, and it needs no GPU.

	Inference job	Training job
Central server	Ray head	FL server
What it does	Holds your endpoint, checks the API key on every request, queues requests and routes each one to a worker	Sends the model to every worker each round, collects what they trained, averages it and sends the result back
Default minimum	2 GB free RAM, 10 GB free disk	1 GB free RAM
A worker drops out	Its replicas move to the remaining workers	The round can’t finish and the job fails
The central server goes down	The endpoint stops; resubmit the job	The job fails; resubmit it

By default a central server is shared, so one machine can coordinate several jobs. For a large model, a full-weight training run or a production endpoint, raise its RAM and disk minimums, ask for a dedicated one in the submit wizard, or order one reserved for you from Support. Sizing for each job type is in Choosing hardware for inference and training; reserving one is in Central servers.

Who it’s for

Teams that want models, not infrastructure.

ML engineers and data teams who already have AI working and want to scale it.
Software teams without ML infrastructure who want a model behind an API without building a serving stack.
Non-technical managers who want AI in their daily workflows and don’t want to know more than the answer their model returns.
Organizations whose data cannot leave: banks, clinics, defense, EU-regulated companies. The platform installs inside your perimeter and runs with the network unplugged.

You don’t need a DevOps hire. If your team can edit a Python file, you have everything the platform asks.

Where the GPUs come from

The same SDK and console work on all three. What changes is who owns the hardware and where the bytes live.

	What it is	When to use it
Public Pool	Our shared GPUs. You name the GPU models you accept and a price ceiling; we only place your job on GPUs priced under it.	Experiments, pilots, workloads without data-residency rules.
Managed Cluster	GPUs reserved for you. Order them in the GPU catalog.	Steady product load and training runs beyond your own capacity.
Private Fabric	The whole platform on your own hardware, inside your perimeter, with no external dependencies.	Data that must not leave the building.

Pick a path

TrainingTrain across many GPUsKeep your training loop, wrap it in one function, and it runs across as many GPUs as you want, in one rack or five buildings.Training overview InferencePaste a model. Get an endpointOpenAI-compatible serving on our GPUs or yours, for any Hugging Face model or any Python you write.Inference overview

Next Install the SDK