The «Your own Docker image» template runs a public image from Docker Hub on a server with a graphics card. The image becomes the server's container with GPU access at once, so you do not write docker run with --gpus yourself. You get SSH and the ports you list at launch.
In short
Your own Docker image runs on cards with 8 GB or more. The most affordable suitable card right now is the RTX 3060 12 GB from 6.95 ₽ per hour, billed by the minute.
Updated 13 October 2026
Memory depends on the image and the program inside it. For inference with 7–8B models 12–16 GB is enough, for example an RTX 4080 16 GB. LoRA training of such models usually needs a 24 GB card, as an RTX 4090 24 GB does. Large models and distributed jobs need 80 GB or more, such as an A100 or H100.
Prices and availability come from the live catalog and update automatically.
If you need a clean system without an image of your own, take the Ubuntu with CUDA template. A ready API server on a GPU without building an image is the vLLM template.
The form has three fields. Each port in the list gets its own protected link on the server page, and SSH is always available.
| Field | What to enter | Example |
|---|---|---|
| Docker image | a public image with a tag from Docker Hub | pytorch/pytorch:2.5.1-cuda12.4-cudnn9-runtime |
| Ports | up to 10 numbers separated by commas, without 22 | 8080, 7860 |
| Disk, GB | at least 10, set at launch and fixed afterwards | 40 |
Check the graphics card inside the container. nvidia-smi shows the card, and the second line checks PyTorch if the image has it.
nvidia-smi
python3 -c "import torch; print(torch.cuda.is_available())"For now the form accepts public images from Docker Hub. A private registry is connected through support: write to @gpunzbot in Telegram, and access will be set up separately.
Billing starts at launch and includes the image download, so keep the image compact. It is better to store large weights on the server's disk rather than in the image, because the image downloads again on every new server. Interruptible servers cost less than standard ones: see the guide interruptible servers.