PyTorch VPS Hosting — from €5.49/mo
Train neural networks on your own server — a PyTorch VPS on NVMe, 5-hour test period, so you risk nothing
Launch PyTorch now
A PyTorch VPS is a virtual server with full root access for training and running deep learning models. You install Linux, Python and the PyTorch stack with the CUDA library versions you need, and the heavy compute runs on the server instead of your laptop. Plenty of RAM and fast NVMe drives keep large datasets and checkpoints close at hand. Plans start at €5.49/mo, with a 5-hour test period — check the training speed on your own data before paying.
Plans
VPS for PyTorch
Standard 1 (4 vCPU, 4 GB) and Standard 2 (4 vCPU, 6 GB) suit code debugging and inference of small models. CPU training scales with cores — take Power 3 (16 vCPU, 24 GB) or higher, and training from scratch needs a GPU configuration. 5-hour test period before payment.
Compare all plansStandard 1
| 4 vCPU · AMD EPYC |
| 4 GB RAM |
| 50 GB NVMe |
| Unlimited traffic |
| 1× IPv4 · KVM |
| 24/7 support |
Standard 2
| 4 vCPU · AMD EPYC |
| 6 GB RAM |
| 60 GB NVMe |
| Unlimited traffic |
| 1× IPv4 · KVM |
| 24/7 support |
Your own PyTorch stack
Install any versions of Python, PyTorch and CUDA libraries — for torchvision, transformers and your models, with no one else's limits.
Compute in the cloud
Training runs on the server around the clock without tying up your computer — start long epochs and leave them running.
Data on NVMe
Fast drives and spare RAM keep large datasets and checkpoints next to the model, speeding up training iterations.
How to run PyTorch on a VPS
Choose a plan and an image
Pick a plan — Standard 1 (4 vCPU, 4 GB) or Standard 2 (4 vCPU, 6 GB) — and select a Linux image, Ubuntu or Debian, for your deep learning stack.
Install the OS yourself
Through the panel you deploy the chosen OS on a KVM server yourself — so the environment stays fully under your control.
Set up PyTorch
Over SSH or the KVM console install Python, PyTorch and the libraries you need, upload the dataset and start training.
A server for your models
We will suggest which plan fits your model and dataset size: RAM volume, vCPU count and NVMe space for data and checkpoints. For GPU workloads we will put together a separate configuration.
We will help set up the environment and secure remote access. Configurations with a graphics card are on the GPU servers page; the alternative framework is a TensorFlow VPS.
Help me size itFrequently asked questions
Didn't find your answer?
Write to us via the contact page.
Any — you install the Python and PyTorch versions you need, with matching CUDA libraries, on your own server.
Count your data and checkpoints: a mid-size model checkpoint is hundreds of megabytes, an image dataset is tens of gigabytes. A starting reserve of 80–120 GB NVMe comes with Standard 3 and Standard 5.
The base plans focus on vCPU, RAM and NVMe — there is no graphics card in them. For GPU workloads we will put together a configuration separately — send us a request.
Yes, CPU mode is fully supported: it suits inference, fine-tuning small models and debugging code. Training from scratch needs a graphics card.
CPU training scales well with cores: 16 vCPU (Power 3, €27.99/mo) compute noticeably faster than eight. Take 24 GB of RAM or more for mid-size datasets.
Yes, root access lets you install any PyTorch build and drivers. On configurations without a graphics card the CPU version is used — it installs normally via pip.
Connect through the KVM console: it works at the hardware level and stays available when the system fails to boot or SSH is closed.
Yes, as your model grows you can move to a higher plan — adding vCPU, RAM and disk space for larger datasets.