10–16 GB
- SDXL and Flux in fp8
- Chat models up to 14B in 4-bit
- Whisper and text to speech
from €0.12/h
GPU rental · billed by the minute · in euros
Rent the VRAM your model needs. Not a gigabyte more.
Most AI jobs are limited by video memory, not by the card’s name. Tell us what you want to run, and we’ll show the cheapest GPU it fits on, from €0.12 an hour.
- 20 GPU modelsfrom 10 to 192 GB
- Per minuteminimum 15 minutes per launch
- 2–5 minutesusually until SSH and Jupyter
- Euros, card or SEPAVAT shown at checkout
What do you want to run?
Needs about 16 GB
Fits on
Rough numbers for a typical setup. Longer context, bigger batches and higher resolution need more.
Choose by memory, not by model number
An RTX 3090 and an RTX 4090 hold the same 24 GB. The second is faster, the first costs half as much. Start from what has to fit, then decide how fast you need it.
24–32 GB
- Qwen-Image and Wan 2.2 video
- Chat models up to 32B in 4-bit
- LoRA training for Flux and SDXL
from €0.30/h
48 GB
- Llama 3.3 70B in 4-bit
- Long video at full resolution
- QLoRA on 30B models
from €0.46/h
80–96 GB
- 70B models with long context
- Full fine-tuning of 7–8B models
- Big-batch inference with vLLM
Cards
from €1.46/h
141–192 GB
- 235B mixture-of-experts in 4-bit
- Training with large batches
- The biggest open models on one card
from €5.43/h
Three ways to use a GPU
Same machines, same per-minute price. Pick the way you like to work.
One-click apps
ComfyUI with Flux and Wan, a private chat on an open model, text to speech. The server installs the app and its models; you open it in the browser.
ComfyUIPrivate chatText to speech
Browse apps → 
Jupyter and SSH
A clean machine with NVIDIA drivers and CUDA 12. Bring any Docker image or start from PyTorch, vLLM or Ollama. Connect from VS Code over SSH.
PyTorchvLLMOllamaDocker
Rent a server → 
API
Start and stop servers from your own code: spin up a GPU for a batch job at night and switch it off when the queue is empty.
RESTAPI keysPer-minute billing
API docs → The meter runs only while the GPU works
You see the price before you launch. Stop the server, and billing stops the same minute. Your files can stay with us for a few cents a month.
- Price per minute is fixed at launch
- Minimum 15 minutes per launch, then minute by minute
- We email you if a server sits idle, and stop it before it eats your balance
- Server started
- ComfyUI ready, Flux loaded
- 112 images rendered
- Stopped, files saved
39 minutes€0.38
billed for 39 minutes, not for an hour
Your models stay planted between sessions
Models, LoRAs and outputs stay in your saved folder while the GPU is off. The next server, even a different card, starts with everything in place, so you don’t download 30 GB again.
€0.03per GB per month
€0.60a month for 20 GB of models
Any cardthe folder moves with you
What people grow on VRAMfarm
Machines come with recent NVIDIA drivers and CUDA 12. Any Docker image works.
Language models
Llama, Qwen, Mistral and DeepSeek with vLLM or Ollama. 24 GB runs a 32B model in 4-bit.
Images and video
ComfyUI with Flux, SDXL, Qwen-Image and Wan 2.2. The template opens the interface in your browser.
Fine-tuning
LoRA and QLoRA on 24–48 GB, full fine-tuning on the A100, H100 and H200.
Rendering and 3D
Blender Cycles, Octane and Redshift. Several machines in parallel when a deadline is close.
Buy a card, use a big cloud, or rent on VRAMfarm
| Buy an RTX 4090 | Big cloud | VRAMfarm | |
|---|---|---|---|
| To start | €1,800 upfront | A contract, sometimes prepayment | from €10 |
| You pay for | Hardware, power, upkeep | By the hour, often with a minimum | By the minute, for actual use |
| Memory | 24 GB, forever | Mostly 80 GB data-centre cards | 10 to 192 GB, 20 models |
| Upgrades | Sell and buy a new one | Wait until the cloud buys new cards | Pick a different card next time |
| Idle time | You always pay | Depends on the plan | Stopped means not paying |
From the blog
Go to the blog →30 guides on choosing a GPU by memory, running models and keeping the bill down.
11 November 2026 How much VRAM do I need? A table for LLMs, Flux, Wan 2.2 video and training
VRAM = weights × bytes per parameter + KV cache + overhead. A 70B model in 4-bit needs about 43 GB of weights. Tables by model, quantisation and card.
11 November 2026 Best GPU for Flux, SDXL and ComfyUI: 16 GB vs RTX 3090, 4090, 5090 and 48 GB cards
SDXL runs on 12 GB, Flux on 16 GB, Qwen-Image and HiDream need 24 GB, Wan 2.2 14B 32–48 GB. Which card to rent for ComfyUI and when the faster one is cheaper.
12 November 2026 Run Llama 70B on one GPU: 4-bit or FP8, 48 GB or 80 GB, and real speeds
Llama 3.3 70B in 4-bit is 40–43 GB: one 48 GB card at 8–16k context, 80 GB for long documents or several users. Formats, commands, tokens/s.
Frequently asked questions
What is VRAMfarm?
An hourly GPU rental service for customers in the EU. Developers, students and companies rent graphics cards for AI, rendering and compute, or launch ready-made apps for images, video and chat. Prices are in euros and you pay by the minute.
How does it work?
The catalogue lists machines with graphics cards in data centres. You choose what to run: a ready-made app for images, video or chat, or a developer server with PyTorch and Jupyter. Everything runs in an isolated container with SSH and Jupyter access, billed by the minute. The server disk is wiped when you stop, but you can save your working folder and get it back on your next server.
Why is it cheaper than the big clouds?
We don’t build our own data centres. We bring available machines from several providers into one catalogue, so an hour costs less than at clouds that run their own facilities.
How do I rent a GPU?
Sign up, confirm your email and top up your balance. Then click Rent a GPU in the dashboard, choose what to run, then the card and the machine, and click Launch. Your balance must cover one hour of the chosen machine. Connection details appear in the dashboard and arrive by email.
How much does GPU rental cost?
From €0.12 an hour for an RTX 3060, €0.30 for an RTX 3090, €0.59 for an RTX 4090, €1.46 for an A100 and €3.90 for an H100. Prices are per hour for a machine with one card, billed by the minute. Current prices for every model are in the server catalogue.
How am I billed?
Every minute a server runs, 1/60 of its hourly price is deducted from your balance. The minimum charge per launch is 15 minutes. To launch, your balance must cover at least one hour of the server; that amount is not charged up front. Internet traffic is billed separately, as used. If your balance runs out, the server stops and we email you.
Is traffic charged?
Yes, per gigabyte downloaded and uploaded. The rate depends on the machine, is usually around a cent per GB or less (B200 and H200 can cost more) and is shown before you launch. We favour machines with cheap traffic.
Is there a minimum rental period?
No long commitment: you pay by the minute, with a minimum charge of 15 minutes per launch. Launch a machine, test your code and stop it.
How do I top up my balance?
In the Balance section of the dashboard, by card or SEPA transfer in euros, from €10. Payments go through a payment provider that we will announce before launch. Until then, top-ups may be unavailable.
What about VAT?
Prices on the site are in euros. How VAT applies to your account and country will be shown at checkout, before you pay.
Your first GPU is a minute away
Sign up with email and password, top up from €10, and launch a server or a ready-made app.